Skin cancer is a dangerous and widespread disease, and early detection increases the survival rate. It is found that a skilled dermatologist usually follows a series of steps, starting with the naked-eye observation of suspected lesions, then dermoscopy followed by biopsy. This would consume time and the disease may advance to later stages. Moreover, accurate diagnosis is subjective, depending on the skill of the clinician. In order to diagnose skin cancer speedily at the earliest stage, we need extensive research solutions by developing computer image analysis algorithms. We have come up with an AI Based Solution Compliance for predictive modelling and diagnosis of Skin Cancer disease.
What it does
~Trained Neural Network to calculate and predict skin cancer alongwith the probability score on a single upload of image.
~Display the percentage of melignant and benign.
~Web based frontend framework UI for user (patient).
~Recommending the nearest doctors to the patient on single click of button.
~Human interacting both text and voice based chatbot to assist anyone who visits the website.
How I built it
We aimed at creating a web-based frontend framework UI for users that allows them to get diagnosed by just uploading their image. Our model is trained on a Neural Network that instantly identifies the disease and recommends the nearest doctor locating the user's location. We used Transfer Learning for this purpose. And the pre-trained model is ResNet50. We here provided a chatbot that is both voice and text controlled(built with dialogflow), to help users in their queries regarding the application, skin cancer precautions, harms, types, etc. Users are directed to nearest dermatologist using their geolocation. We used Map here API to carry out hospital's information.
Challenges I ran into
The model was initially trained out to classify disease as malignant or benign. And if some user uploaded a random image that is not skin type, it was unable to recognize it, since it detected disease pixel by pixel. So we overcame this problem by adding third classification to identify not a skin cancer image.
Accomplishments that I'm proud of
Accuracy of the trained model is pretty awesome -85%. Transfer Learning reduced the training time and increased the accuracy.
What I learned
Searching nearest doctor and extracting right indices for hospital information from Here Map API and displaying it along with map. We also came up with hosting the website with flask.
What's next for Skin cancer Sentinel with doctor recommendation
~Our future goal is to modify this web app into an application that keeps the track of patients maintaining dashboards.
~This will allow users and doctors to store all related information and reports on the application itself so that patients don't have to carry their reports everytime.
~Video consultancy and discussion forums can be maintained on the app for better insights.
~Comprehensive Profile Of Doctors & Hospitals will enable users to explore about hospital and dermatologists they want to seek out.
Log in or sign up for Devpost to join the conversation.