AU2021105721A4

AU2021105721A4 - System and method for classifying skin cancer lesions using deep neural networks and transfer learning

Info

Publication number: AU2021105721A4
Application number: AU2021105721A
Authority: AU
Inventors: Danny Joel Devarapalli; Harshit Gorijavolu; Venkata Sai Dheeraj Mavilla; Sri Anjaneya Nimmakuri; Sai Prashanth Reddy Karri
Original assignee: Individual
Current assignee: Individual
Priority date: 2021-08-17
Filing date: 2021-08-17
Publication date: 2021-11-18
Anticipated expiration: 2029-08-17

Abstract

The present invention generally relates to a system for classification of skin cancer lesions using deep neural networks and transfer learning comprises a data acquisition unit configured for collecting dermoscopic images, precancerous and cancerous images, which covers all characteristics of each type of skin cancer; a pre-processing unit for removing unwanted images and cropping images thereby verifying each and every image and dividing images of each folder into three original, edit, crop segments; an image augmentation unit for generating images to divide equal number of images to each class; and a classification unit equipped with a set of classification approaches for classification of skin cancer lesions using deep neural networks and transfer learning for detecting cancer and type of cancer. N (D 03 -F Li n VC Lij 0 l3 f 4-J 4 *L F M itL .~i rj 03 j 0t LP* ~ . ~JL.J ~ fa fiai f1 0 Wr~ i2 . 02 Li fa (Nw

Description

N (D

03 -F

Li

n VC

Lij

0 l3 f

itL .~ij 4-J 4

*L F M

rj 03 0t LP* ~ .

~JL.J ~ fa

fiai f1 Wr~i2 0 .

02 Li fa (Nw

SYSTEM AND METHOD FOR CLASSIFYING SKIN CANCER LESIONS USING DEEP NEURAL NETWORKS AND TRANSFER LEARNING FIELDOFTHE INVENTION

The present disclosure relates to a system and method for classifying skin cancer lesions using deep neural networks and transfer learning.

BACKGROUND OF THE INVENTION

There are many different forms of skin illnesses and malignancies, some of which are deadly. Skin cancer must be treated at an early stage or it will be fatal. Many individuals overlook skin lesions, which can be fatal. In rural regions, when hospitals are understaffed, it is difficult to diagnose illnesses in their early stages.

Currently, the only diagnostic method accessible is a series of lengthy laboratory investigations. However, a prediction system is required to allow a patient or dermatologist to quickly determine if a lesion is malignant or benign, allowing cancer to be treated in its early stages. Despite the fact that early identification of Melanoma might significantly increase survival rates, almost tens of thousands of individuals die each year. People frequently ignore their health problems in the early stages and rush to the hospital when they become worse. This can be prevented by assisting individuals in comprehending or diagnosing the illness. The majority of the time, skin cancers do not cause discomfort, but they are easily apparent since cancer is nothing more than the abnormal development of skin cells.

In the view of the forgoing discussion, it is clearly portrayed that there is a need to have a system and method for classifying skin cancer lesions using deep neural networks and transfer learning.

SUMMARY OF THE INVENTION

The present disclosure seeks to provide a system and method to determine class of a skin cancer lesion to its correct classified class of skin cancer using a deep convolutional neural network and other segmentation methods.

In an embodiment, a system for classifying skin cancer lesions using deep neural networks and transfer learning is disclosed. The system includes a data acquisition unit configured for collecting dermoscopic images, precancerous and cancerous images, which covers all characteristics of each type of skin cancer. The system further includes a pre-processing unit for removing unwanted images and cropping images thereby verifying each and every image and dividing images of each folder into three original, edit, crop segments. The system further includes an image augmentation unit for generating images to divide equal number of images to each class. The system further includes a classification unit equipped with a set of classification approaches for classification of skin cancer lesions using deep neural networks and transfer learning for detecting cancer and type of cancer.

In an embodiment, original, edit, crop segments are stored as in target dataset, wherein some images contained watermarks and clinical markings that is avoided or edited using a tool GIMP.

In an embodiment, image augmentation employed to artificially expand the dataset, wherein image augmentation parameters that are used to increase data sample count are zoom, shear, rotation, preprocessing function and the like that results in generation of images having these attributes during training of Deep Learning models.

In an embodiment, basal cell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma.

In an embodiment, deep convolutional neural network, transfer learning comprises VGG16, InceptionResNetV2 and ResNet101 architectures are trained with different learning rates, epochs and batch sizes.

In an embodiment, feature extraction in deep convolutional neural network is performed by alternating convolution layers with subsampling layers, wherein classification is performed with dense layers followed by a final SoftMax layer.

In an embodiment, ReLu activation is engaged with deep convolutional neural network for every layer, SoftMax in the end, and a drop out of 0.4(40%) to generalize data, thereby avoiding overfitting.

In an embodiment, input images in transfer learning are extracted and resized to 224x224x3, wherein a GlobalAveragePooling2D Layer, Dropout of 0.5 for VGG16 and ResNet101 and dropout of 0.4 for InceptionResNetV2 is added with activation ReLu and a Dense Layer of 3, for 3 classes in the Dataset.

In an embodiment, Adam optimizer with a user defined learning rate is employed to use a smaller learning rate during transfer learning.

In an embodiment, a method for classification of skin cancer lesions using deep neural networks and transfer learning is disclosed. The method includes collecting dermoscopic images, precancerous and cancerous images, which covers all characteristics of each type of skin cancer. The method further includes removing unwanted images and cropping images thereby verifying each and every image and dividing images of each folder into three original, edit, crop segments. The method further includes generating images by employing an image augmentation unit to divide equal number of images to each class. The method further includes classifying skin cancer lesions using deep neural networks and transfer learning for detecting cancer and type of cancer using a classification unit equipped with a set of classification approaches.

An object of the present disclosure is to provide a user-friendly environment for detecting cancer and type of cancer using a user device.

Another object of the present disclosure is to reduce melanoma mortality.

Yet another object of the present invention is to deliver an expeditious and cost-effective method for classifying skin cancer lesions using deep neural networks and transfer learning.

To further clarify advantages and features of the present disclosure, a more particular description of the invention will be rendered by reference to specific embodiments thereof, which is illustrated in the appended drawings. It is appreciated that these drawings depict only typical embodiments of the invention and are therefore not to be considered limiting of its scope. The invention will be described and explained with additional specificity and detail with the accompanying drawings.

BRIEF DESCRIPTIONOF FIGURES

These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

Figure 1 illustrates a block diagram of a system for classifying skin cancer lesions using deep neural networks and transfer learning in accordance with an embodiment of the present disclosure; Figure 2 illustrates a flow chart of a method for classifying skin cancer lesions using deep neural networks and transfer learning in accordance with an embodiment of the present disclosure; Figure 3 illustrates exemplary profiles of BCC, Melanoma, SCC images in accordance with an embodiment of the present disclosure; Figure 4A and 4B illustrates augmentations of brightness and rotation in accordance with an embodiment of the present disclosure; Figure 5A, 5B, 5C, 5D, and 5E illustrates learning curves in accordance with an embodiment of the present disclosure; Figure 6A and 6B illustrates quadrant 3 strategy in accordance with an embodiment of the present disclosure; and Figure 7A, 7B, 7C, and 7D illustrates learning curves in accordance with an embodiment of the present disclosure.

Further, skilled artisans will appreciate that elements in the drawings are illustrated for simplicity and may not have necessarily been drawn to scale. For example, the flow charts illustrate the method in terms of the most prominent steps involved to help to improve understanding of aspects of the present disclosure. Furthermore, in terms of the construction of the device, one or more components of the device may have been represented in the drawings by conventional symbols, and the drawings may show only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having benefit of the description herein.

DETAILED DESCRIPTION:

For the purpose of promoting an understanding of the principles of the invention, reference will now be made to the embodiment illustrated in the drawings and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the invention is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the invention as illustrated therein being contemplated as would normally occur to one skilled in the art to which the invention relates.

It will be understood by those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not intended to be restrictive thereof.

Reference throughout this specification to "an aspect", "another aspect" or similar language means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. Thus, appearances of the phrase "in an embodiment", "in another embodiment" and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.

The terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such process or method. Similarly, one or more devices or sub-systems or

F; elements or structures or components proceeded by "comprises...a" does not, without more constraints, preclude the existence of other devices or other sub-systems or other elements or other structures or other components or additional devices or additional sub-systems or additional elements or additional structures or additional components.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The system, methods, and examples provided herein are illustrative only and not intended to be limiting.

Embodiments of the present disclosure will be described below in detail with reference to the accompanying drawings.

Referring to Figure 1, illustrates a block diagram of a system for classifying skin cancer lesions using deep neural networks and transfer learning is illustrated in accordance with an embodiment of the present disclosure. The system 100 includes a data acquisition unit 102 configured for collecting dermoscopic images, precancerous and cancerous images, which covers all characteristics of each type of skin cancer.

In an embodiment, a pre-processing unit 104 is in continuation with the data acquisition unit 102 for removing unwanted images and cropping images thereby verifying each and every image and dividing images of each folder into three original, edit, crop segments.

In an embodiment, an image augmentation unit 106 is connected with the pre-processing unit 104 for generating images to divide equal number of images to each class.

In an embodiment, a classification unit 108 is equipped with a set of classification approaches for classification of skin cancer lesions using deep neural networks and transfer learning for detecting cancer and type of cancer.

R

Figure 2 illustrates a flow chart of a method for classifying skin cancer lesions using deep neural networks and transfer learning in accordance with an embodiment of the present disclosure. At step 202, the method 200 includes collecting dermoscopic images, precancerous and cancerous images, which covers all characteristics of each type of skin cancer.

At step 204, the method 200 includes removing unwanted images and cropping images thereby verifying each and every image and dividing images of each folder into three original, edit, crop segments.

At step 206, the method 200 includes generating images by employing an image augmentation unit 106 to divide equal number of images to each class.

At step 208, the method 200 includes classifying skin cancer lesions using deep neural networks and transfer learning for detecting cancer and type of cancer using a classification unit 108 equipped with a set of classification approaches.

Figure 3 illustrates exemplary profiles of BCC, Melanoma, SCC images in accordance with an embodiment of the present disclosure. The data for testing the system is collected from various sources including

q

HAM 10000 dataset, ISIC Dataset, Web scraping, social media, other sites. Dermoscopic images, Precancerous and Cancerous images, are collected which covers all the characteristics of each type of skin cancer. Dermoscopic images are collected for the hospitals and dermatologists by which the predictor system can help them better in identifying the disease. The data collected contained clinical images. Deep learning networks need thousands of images to learn, to give better results, downloading thousands of images manually is difficult, Selenium a portable framework is used for testing web applications to scrape images from the web added them to each class folder accordingly and renamed all images using a simple python script.

The images required for the system are BCC, Melanoma, SCC images that are filtered from the dataset of 10015 dermatoscopic images. The required images are filtered with the help of metadata provided along with the HAM10000 dataset.

Images are also collected from Instagram by using hashtags, only the necessary images focusing the skin lesions are cropped and edited to add these images to the dataset. The unnecessary images are deleted from the dataset.

587 BCC, 603 Melanoma, 225 SCC images are collected from an open-source public access archive of skin images to test and validate the proposed standards.

Data cleaning include eliminating unnecessary pictures and cropping the photographs to focus on the lesion. The pictures are posted to Drive in several folders, such as HAM 10000, Web, and social media, after data gathering. Each and every image is manually checked and the images of each folder are divided into three segments, namely, Original, Edit, Crop. Original - The images which require no changes. Edit - The images which require watermark, markings removal. Crop - The images which are to be

1i cropped to focus the lesion. The pictures are altered as described above, and they are saved as our target dataset. Some photographs included watermarks and clinical markings on lesions that might have confused the model; these images are either removed or modified with the GIMP tool.

The completed picture data is ready to be formed into a Deep Neural Network in this stage. Squamous and Basal Cell Carcinoma images are found because our dataset is imbalanced, which means our classes have different numbers of images. Many images for melanoma are obtained because the data available is huge because it is a very common cancer, but fewer images are found for Squamous and Basal Cell Carcinoma because our dataset is imbalanced, which means our classes have different numbers of images. When data is unbalanced by class, the model is more likely to overfit to the class with the most classes, or to be sensitive to just the class with the most pictures. The ImageDataGenerator class is used, which creates pictures with given image augmentations, to prevent this susceptibility to overfitting.

Figure 4A and 4B illustrates augmentations of brightness and rotation in accordance with an embodiment of the present disclosure. Image augmentation is a technique for artificially increasing the size of a data set. This is useful for a data set with a small number of data samples. In the case of Deep Learning, this is a poor condition since the model tends to over-fit when only a few data samples are used to train it. Zoom, shear, rotation, preprocessing function, and other image augmentation parameters are commonly employed to enhance the data sample count. When these parameters are used, pictures with these properties are generated during Deep Learning model training. In general, picture samples created by image augmentation result in an approximately 3x to 4x increase in the number of current data samples. Custom parameter values are utilized, but only those that are useful to the data system in question; those that would alter the picture too much by eliminating lesion features/information are not used. Using this tool from Keras, 3552 images are generated for each class, and total dataset image files to 10,656 images. The parameters and their respected values are brightness range=[0.3,1.0], zoom range=[0.5,1.0], horizontal flip= True, rotation range = 90, and vertical flip = True.

Figure 5A, 5B, 5C, 5D, and 5E illustrates learning curves in accordance with an embodiment of the present disclosure. Model training for deep learning is the most time-consuming task, thus the system is trained on Colab, which lets customers to utilize their Computer Engine Backend integrated GPUs and TPUs for free, along with over 12GB of RAM. To train on our data, a basic CNN architecture is built initially.

Feature extraction and categorization are the first steps of a standard CNN architecture. Convolution layers and subsampling layers are alternated for feature extraction. Dense layers are used for classification, followed by a SoftMax layer. This design outperforms a fully connected feed forward neural network when it comes to picture categorization. A basic CNN architecture contains the following.

* Filters is the number of desired feature maps.

* Kernel size is the size of the convolution kernel. A single number 5 means a 5x5 convolution.

* Padding is either 'same' or 'valid'. Leaving this blank results in padding='valid'. If padding is 'valid' then the size of the new layer maps is reduced by kernelsize-1. For example, if you perform a 5x5 convolution on a 28x28 image (map) with padding='valid', then the next layer has maps of size 24x24. If padding is 'same', then the size isn't reduced.

* Activation is applied during forward propagation. Leaving this blank results in no activation.

* 'ReLu' activation is used for every layer and SoftMax in the end, a drop out of 0.4(40%) is used to generalize data, thereby avoiding overfitting. Our architecture takes in an input size of (28,28,1) 'grayscale' image and consists of

* Two Convolutional Layers with feature map 32x32 and kernel size 3x3, with activation 'ReLu' and one Convolutional Layer with feature map 32x32, kernalsize 5x5 with stride 2.

* Two Convolutional Layers with feature map 64x64 and kernel size 3x3, with activation 'ReLu' and one Convolutional Layer with feature map 64x64, kernalsize 5x5 with stride 2.

* Flatten layer followed by a fully connected layer, dense - 128, a dropout of 0.4, dense -3(Number of classes present).

* The following model is trained with batch size = 32 and epochs= 100, gave us training accuracy of 0.97 and on validation gave 0.82.4. The learning curves are shown below in Figure 5A and Figure 5B.

The curves show that the model is overfitted since it is trained too successfully on data. Validation accuracy has its ups and downs. After then, it is attempted to alter the architecture slightly. After each layer, Batch Normalization is applied, as well as two dropouts, one with 0.4 after two layers and the other with 0.5 before the last completely linked layer. Kernel regularizer=12(0.001), bias regularized=12(0.001) is used in this application. After 100 epochs, the accuracy is 0.8647, and after training, it is 0.9947. The curves observed smoothed down, although this is not the best model. ILSVRC Models that are trained on the ImageNet dataset are used to train the system. It is utilized VGG16, InceptionResNetV2, and ResNet101. The data is trained without ImageNet weights and simply on their architectures to examine the results, and then the model is fine tuned and trained with ImageNet pretrained weights using the transfer learning technique.

ResNet101, VGG, and InceptionResNetV2 are employed with and without transfer learning, with weights set to None and weights set to 'ImageNet' with transfer learning. To prevent extracting data every time the system runs, a pickle object file with picture data ready is stored in the disc.

Keras applications are used to employ the VGG16, InceptionResNetV2 and ResNet101 architectures, with the weights option set to None in the documentation supplied by Keras/applications. Because most ImageNet models utilize this image format and contain top = False, the input images are extracted and scaled to 224x224x3. A GlobalAveragePooling2D Layer is added to this, as well as a Dropout of 0.5 for VGG16 and ResNet101, and a Dropout of 0.4 for InceptionResNetV2, both with the activation 'ReLu' and a Dense Layer of 3 for the Dataset's three classes. Since each image corresponds to just one kind of cancer, the loss function for all models is 'sparse categorical cross entropy', and the metrics values are all equal to ['accuracy'].

A Residual neural network (ResNet) is an artificial neural network (ANN) of a kind that builds on constructs known from pyramidal cells in the cerebral cortex. Residual neural networks do this by utilizing skip connections, or shortcuts to jump over some layers. ResNet-N is a deep residual network that is N layers deep. It is a subclass of convolutional neural networks, with ResNet most popularly used for image classification.

• Stochastic gradient descent is used with learning rate of 0.01, momentum of 0.9.

* With Reset's 101 layers, the Trainable params: 42,558,979 , non trainable params: 105,344.

* TRAINING ACCURACY = 0.891 and VALIDATION ACCURACY= 0.8116 , which is an acceptable result, but the difference must be of a smaller value to be called an optimal model. The Learning curves (in Figure 5C) also show the evidence of a little overfitting.

VGG16 comprises an Adam optimizer with a learning rate of 0.0001. With VGG16's Architecture, Total params: 3,588,003, Trainable params:

3,588,003, non-trainable params: 0. The system is trained with Colab with Compute Engine backend integrated GPU. With batch size of 32, epochs 100, and validation slipt 0.2.

* TRAINING ACCURACY = 0.9981 and VALIDATION ACCURACY= 0.8918 , which is an acceptable result, but the difference must be of a smaller value to be called an optimal model. The Learning curves (in Figure 5D) also show the evidence of overfitting. With the difference being more than InceptionV2's.

Inception Layer is a combination of all those layers namely, 1x1 Convolutional layer, 3x3 Convolutional layer, 5x5 Convolutional layer with their output filter banks concatenated into a single output vector forming the input of the next stage.

* Adam optimizer is used with a learning rate of 0.0001.

* With InceptionV2's Architecture, Total params: 54,451,939, Trainable params: 54,391,395, non-trainable params: 60,544. With batch size of 32, epochs 10, and validation-slipt 0.2

* TRAINING ACCURACY = 0.9575 and VALIDATION ACCURACY=

0.8661 , which is an acceptable result, but the difference must be of a smaller value to be called an optimal model.

* The Learning curves (in Figure 5E) also show the evidence of overfitting. With the difference being more.

19;

Figure 6A and 6B illustrates quadrant 3 strategy in accordance with an embodiment of the present disclosure. With transfer learning, instead of starting the learning process from scratch, patterns are started that have been learned when solving a different problem.

This way system leverage previous learnings and avoid starting from scratch. There are basically three types of transfer learning includes train the entire model, train some layers and leave the others frozen, and freeze the convolutional base.

Out of these 3 types, the second way is selected, this strategy is with a small dataset, as 10k images is small compared to the 14million images of ImageNet. It has been observed that there are no similarities between ImageNet data and our dataset, so it is concluded that the dataset is dissimilar to that of ImageNet, taking into consideration the size, Quadrant 3 strategy (from Figure 6A and Figure 6B) suits best to the problem. Keras applications is used to use VGG16, InceptionResNetV2 and ResNet101 architectures, from the documentation provided by Keras/applications, the weights parameter as 'ImageNet'. The input images are extracted and resized to 224x224x3, since most of the ImageNet models use this image format and include top = False. All the models have loss function as 'sparsecategorical crossentropy' since, one image belongs to only one type of cancer, and the metrics values=

['accuracy'] for all.

Figure 7A, 7B, 7C, and 7D illustrates learning curves in accordance with an embodiment of the present disclosure. A Convolutional Layer is added to this with feature maps 64, kernalsize = (3,3), a MaxPooling2D with pool size = 2, a Flatten Layer, a fully connected layer of 256, a Dropout of 0.5 and a fully connected layer 3 with a 'SoftMax' activation.

16F

Adam optimizer is sued with learning rate of 0.0001, to the notion that it's better to use a smaller learning rate during transfer learning. With VGG16's Architecture, Total params: 12,278,915, Trainable params: 12,278,915, non-trainable params: 0. The system is trained with Colab with Compute Engine backend integrated GPU. With batch size of 32, epochs 100, and validation-slipt 0.2. TRAINING ACCURACY = 0.9785 and VALIDATION ACCURACY = 0.9009 , which is an acceptable result, but the difference value can be accepted but system cannot certainly call it an optimal model. The Learning curves(in Figure 7A) also do not show much of the evidence of overfitting. Compared with VGG16 without pre-trained weight.

A Flatten layer is added to the loaded model, a Dropout of 0.4 and finally a Dense layer 3 with 'SoftMax' activation. Adam optimizer is used with learning rate of 0.0001. TRAINING ACCURACY = 0.9927 and VALIDATION ACCURACY = 0.9314 , which is an acceptable result, but the difference value can be accepted but system cannot certainly call it an optimal model. This has been the best result so far, and compared with the InceptionV2 model without using any pre-trained weights, it is obtained that the transfer learning method outperformed. The Learning curves (in Figure 7B) also do not show much of the evidence of overfitting.

With ResNet101's Architecture, Total params: 42,664,323, Trainable params: 42,558,979, non-trainable params: 105,344. With batch size of 32, epochs 100, and validation-slipt 0.2. TRAINING ACCURACY = 0.9992 and VALIDATION ACCURACY = 0.9563 , which is an acceptable result and almost an optimal model, but the difference value is accepted. This has been the best result so far with training and testing accuracies having very less difference, which means that the model has trained well on the training data of 8516 images, generalized well without overfitting and it can 95.63% accurately predicts to new data, and compared with the ResNet101 model without using any pre-trained weights, it can be observed that the transfer learning method outperformed the traditional approaches. The Learning curves (in Figure 7C and Figure 7D) also do not show much of the evidence of overfitting. Compared with ResNet1O1 without pre-trained weights.

(224,224,3) 224x224 pixel with colored image(RGB), and assign each output to a result variable. 3 count variables are defined for each class and using simple for loops, the classes is checked which are predicted by each model, if the modell gives class 1, then county will be incremented, the final maximum value would be printed, like an ensemble method of voting and getting the max vote value. This model is deployed as a static web application using TensorFlow.js which will allow us to load python written models in JavaScript. Using Node.js and Express server to run the scripts in the client browser.

Several architectures are trained with different learning rates, epochs and batch sizes, however ResNet1O1 architecture with ImageNet weights has given us the best accuracy to identify the type of a skin cancer lesion ever to be published or recorded, which is 95.63% with training accuracy of 99.92%, model overfitting is not obtained in this case. Also an ensemble approach which has been said to give better results is implemented by using basic voting mechanism which is written in Python.

The drawings and the forgoing description give examples of embodiments. Those skilled in the art will appreciate that one or more of the described elements may well be combined into a single functional element. Alternatively, certain elements may be split into multiple functional elements. Elements from one embodiment may be added to another embodiment. For example, orders of processes described herein may be changed and are not limited to the manner described herein. Moreover, the actions of any flow diagram need not be implemented in

1R the order shown; nor do all of the acts necessarily need to be performed. Also, those acts that are not dependent on other acts may be performed in parallel with the other acts. The scope of embodiments is by no means limited by these specific examples. Numerous variations, whether explicitly given in the specification or not, such as differences in structure, dimension, and use of material, are possible. The scope of embodiments is at least as broad as given by the following claims.

Benefits, other advantages, and solutions to problems have been described above with regard to specific embodiments. However, the benefits, advantages, solutions to problems, and any component(s) that may cause any benefit, advantage, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential feature or component of any or all the claims.

1q

Claims

WE CLAIM

1. A system for classification of skin cancer lesions u sing deep neural networks and transfer learning, said system comprises:

a data acquisition unit configured for collecting dermoscopic images, precancerous and cancerous images, which covers a II characteristics of each type of skin cancer; a pre -processing unit for removing u nwanted images a nd c ropping images thereby verifying each and every image a nd dividing images of each folder into three original, edit, crop segments; an image augmentation unit for generating images to divide equal number of images to each class; and a classification unit equipped with a set of classification approaches for classification of ski n cancer lesions using deep n eural networks and transfer learning for detecting cancer and type of cancer.

2. The s ystem as claimed in cla im 1 , wherein original, edit, crop segments are stored as in target dataset, wherein some images contained watermarks and clin ical markings that is avoided or edited using a t ool GIMP.

3. The system as cl aimed in claim 1, w herein image augmentation employed to artificially expand said dataset, wherein image augmentation parameters that are used to increase data sample count are zoom, shear, rotation, preprocessing function and the like that results in generation of images having these attributes during training of Deep Learning models.

4. The system as c laimed in claim 1, wherein basa I c ell carcinoma (BCC), squamous cell carcinoma (SCC) and melanoma.

Mo

5. The system as cla imed in cla im 1 , wherei n deep convolutional neural network, transfer learning comprises VGG16, Incep tionResNetV2 and ResN et1O1 arch itectures are train ed with differe nt learning rates epochs and batch sizes.

6. The system as claimed in claim 5, w herein feature extraction in deep convolutional neural network is performed by alternating convolution layers w ith su bsampling laye rs, wherein classification is perf ormed with dense layers followed by a final SoftMax layer.

7. The system as claimed in cla im 6, w herein ReLu activation is engaged with deep convolutional neural network for every layer, SoftMax in t he end, and a drop out of 0.4(40 %) to ge neralize data , thereby avoiding overfitting.

8. The system as claimed in claim 5, wherein input images in transfer learning are extracted and resized t o 22 4x224x3, wh erein a GlobalAveragePooling2D Layer, Dropout of 0.5 for VGG16 and ResNet101 and dropout of 0.4 for Inc eptionResNetV2 is add ed with activation ReLu and a Dense Layer of 3, for 3 classes in said Dataset.

9. The system as claimed in claim 5, wherein Adam optimizer w ith a user defined learn ing rate is employed to u se a smaller learning rate during transfer learning.

10. A method for classification of skin cancer lesions using deep neural networks and transfer learning, said method comprises:

collecting dermoscopic images, precancerous and cancerous images, which covers all characteristics of each type of skin cancer; removing unwanted images and cropping images thereby verifying each a nd every image and dividing images of eac h fold er into thre e original, edit, crop segments; generating images b y employi ng an image augmentation unit to divide equal number of images to each class; and classifying ski n cancer lesions us ing deep neural n etworks and transfer learnin g fo r detecting cancer and type of cancer using a classification unit equipped with a set of classification approaches.

Figure 2

Figure 3

Figure 4A

Figure 4B

Figure 5A

Figure 5B

Figure 5C

Figure 5D

Figure 5E

Figure 6A Figure 6B

Figure 7A Figure 7B

Figure 7C Figure 7D