Multiclass Non-Invasive Malignant Cells Screening using Ensemble of Deep Neural Networks

Project Title

Multiclass Non-Invasive Malignant Cells Screening using Ensemble of Deep Neural Networks

Project Area of Specialization Artificial IntelligenceProject Summary

Skin cancer is known as one of the most hazardous forms of Cancer found in Humans. Skin cancer is a serious disease that requires early detection to improve survival rates. The factors contributing to skin cancer include prolonged exposure to direct ultraviolet (UV) rays and the presence of atypical moles, skin types, and also family medical history. Basal cell carcinoma and squamous cell carcinoma are probably the most common non-melanoma skin cancers. Traditionally, the biopsy method is used for diagnosing and detecting melanoma. This procedure can be excruciatingly painful and time-consuming. For testing purposes, it will take a lot more time.

Early diagnosis drove out various complications caused by skin tumors. Computer-aided automatic diagnostics systems are developed using a variety of methods. Mostly data reduction techniques with controlled threshold and RGB histogram features with region growing segmentation techniques are utilized. The data was largely acquired from hospitals, but the sample size was insufficient for training a machine learning model.

Our proposed ‘Multiclass Noninvasive Malignant Cells Screening System using Ensemble of Deep Neural Networks’ system classifies melanoma and non-melanoma skin cancer. The input for the system is the image of the skin lesion which is suspected to be a melanoma or non-melanoma lesion. Transfer Learning models such as MobileNet, ResnetV2, and DenseNet are used for better edge detection in images with pre-calculated weights. To overcome the overfitting problem, data augmentation (Rotate, Flip, Zoom, MotionBlur) techniques are utilized at small data before training and utilize augmentation pipeline to deal with class imbalance. Skin cancer is highly curable if it gets identified at the early stages. It offers a simple web interface for public release. The user or dermatologist can upload the patient's demographic information along with the skin lesion image, and the model will analyze the data and return results in a fraction of a second. Keeping the larger demographic of people in mind, the basic infographic page develops, which provides a simplistic overview of skin cancer as well as for instructions for using the online tool to obtain the results. The goal of the research is to create an automated classification system for skin cancer using images of skin lesions that are based on image processing techniques.

Project Objectives

1. Our proposed framework will provide a platform for different users can check skin cancer type via skin lesion image and decide on a diagnosis and also provide relevant information about melanoma and other skin cancers to people.
2. Our proposed system will facilitate both patients and dermatologists because it is much easier to get an overview of the skin cancer cell by uploading information such as age, gender, localization, and skin lesion image to decide which skin specialist user needs to visit or if it is normal.
3. The proposed system allows for the reduction of time-consuming procedures that cause additional complications.
4. Our suggested system facilitates low-cost detection methods, such as AI could change the patient’s diagnostic pathways, allowing greater effectiveness during the health care service.

Project Implementation Method

1. We’re working on a web-based platform that’s particularly a healthcare diagnosis system that classifies different malignant skin cells type based on information such as age, sex, localization, a skin lesion image, etc.
2. For training of our proposed system, data was gathered from Dataverse.harvard.edu and Kaggle. Features contained in the dataset include skin lesion images and metadata.
3. At first, the User will register at our platform using his/her email address and phone number, which will be later saved in the PhpMyAdmin database. Then He/she will sign into our platform using his/her credentials and provide basic information and upload a skin lesion image.
4. Using technology, we are developing a web utility that will connect concerned individuals and medical institutions into a platform that helps them to classify and diagnose malignant cells.This platform will include information such as:
   1. Patient’s age

   2. Patient’s gender

   3. Patient’s skin lesion image

   4. Patient’s localization

5. keeping in mind different demography, our suggested system will provide an interactive webpage with information about melanoma and non-melanoma skin diseases and how to use the online suggested platform to get the results.
6. Different transfer learning models can be utilized such as ResnetV2, MobileNet, and DenseNet for better image processing which will improve the validation and accuracy then integrate our deep learning model into a web platform that will be easily utilized by users.
7. Users can also check via skin lesion image what type of skin specialist he/she needs to visit or whether it's completely normal.
8. Our main goal is to save time and medical resources. With our suggested platform, users/dermatologists can decide whether a biopsy is necessary or not that saving a lot of medical resources and time.
9. GUI for user interaction will be developed using Django/Flask. HTML, CSS3 and bootstrap, and JavaScript are used for the front-end interactive user interface.

Benefits of the Project

1. Our Proposed web-based framework will be available to different users where it’ll help them to classify a patient’s malignant skin cells and to carry out a diagnosis.
2. The proposed framework saves a significant number of medical supplies and resources that would otherwise be used in time-consuming biopsies.
3. Medical supplies are limited in developing countries. The integration of our system into their healthcare system has some long-term benefits.
4. Our suggested system can help dermatologist to detect malignant skin cell in early stages which reduces the chances of getting serious outcome.
5. Our system provides more efficient and comfortable way than biopsy method for both patients and doctors because it is a timeless and painless method.

Technical Details of Final Deliverable

The final deliverable will be a web-based framework that will help users to detect and classify malignant skin cells type. It’ll take an image as input and return a probability prediction of is it a malignant skin cell? If yes then which type does it belongs to.

There will be only one type of account, which is a user account. That account must have all necessary privileges.

The user will create an account for him/herself using an email address or phone number, which will be saved in the phpMyAdmin database. After signing into his account, there’s an interactive GUI where different utilities can be shown. Users can select to see Information about malignant diseases, disease classification, and how to use the online tool pages.

At the classification phase, the system will need some parameters such as age, sex, gender, etc. After providing information about him/herself, he/she will upload a skin lesion image that is suspected to be malignant, which will be classified by our deep CNN transfer learning-based model (DCNNTL), which will use in our system for image classification and identification to extract useful patterns.

After identifying the useful patterns in the indented skin lesion image, our system will perform some calculations based on the pre-trained model and provide an estimated prediction that will display to screen that the user might suffer from any malignant disease or it’s completely normal.

Our proposed platform aims to aid in cancer cell classification while also saving time and medical resources that would otherwise be spent on naked-eye examinations and time-consuming biopsies.

Final Deliverable of the Project Software SystemCore Industry ITOther Industries Health Core Technology Artificial Intelligence(AI)Other Technologies OthersSustainable Development Goals Good Health and Well-Being for PeopleRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	9500
Google Colab Pro	Equipment	3	2000	6000
Printing & binding	Miscellaneous	2	1200	2400
DVDs & DVD writing	Miscellaneous	2	300	600
Stationary	Miscellaneous	1	500	500