Skin Cancer Detection

Project description

The goal of this project is to evaluate the impact Deep Learning would make on the skin cancer detection process. The main aim of this project is to evaluate how well AI detection using polygon labels works. Aside from this, there are several other research questions that can be answered such as edge detection and classification of specific types of cancers.

The goal of this project is not to have a fully functioning application, but to have one or more optimized models that are evaluated on how well they perform.

Main research question: How can deep learning improve skin cancer detection using histological microscopic images?

Context

After a skin cancer surgical procedure, a histopathology biopsy image is used to verify whether all of the cancer has been removed during the procedure. The evaluation of this image is a lengthy, manual procedure. During this time period, patients stay at the clinic as the results determines whether they need to return to surgery or are allowed to recover at home. MOHSA wants to accelerate the bulk of the work, making the process faster but keeping the expertise and decision making of dermatologists.

Results

Model Development and Evaluation:
- Conducted experiments with the YOLO model, addressing data challenges and optimizing results.
- Completed data pre-processing, image resizing, and initial runs with the DETR model. Finished format conversion and received debugging assistance.

Data Preparation and Hosting:
- Prepared two datasets (image-mask and image-bounding box) for hosting on Hugging Face.
- Emphasized the importance of well-documented decision-making notebooks for knowledge transfer.
- Consolidated data processing steps (tiling, splitting) into a single, well-commented notebook for improved collaboration and future development.

Validation and Technological Research Level:
- Actively tested and refined YOLO and DETR models, focusing on achieving satisfactory performance.
- Progressed in data preparation for various project stages, including datasets for Hugging Face.
- Committed to well-documented notebooks for future knowledge transfer.

These achievements demonstrate progress in adapting established deep learning models (YOLO, DETR) to the specific requirements of medical image data, aligning with current technological research in medical image analysis.

About the project group

We are a team of four Advanced AI students with passion for AI and Data Science.

Video