ML Projects

Project 1

Skin-lesion Image Analysis using Machine Learning

This project was developed during my Fall 2025 for self-study project under Prof Yao Wang during NYU MSEE program.

Objectives

This research project focuses on exploring the recent advancement in computer-aided diagnosis (CAD) pipeline to triage melanoma risk from lesion images and associated metadata, targeting a pre-biopsy decision support setting where low specificity i.e. True Negatives Rate are clinically costly for biopsy confirmation.

Results

The work is implemented around a challenge that emphasizes (i) extreme class imbalance between the benign and malignant class, (ii) generalization to non-dermoscopic imaging quality, and (iii) sensitivity-focused evaluation using partial AUC above a minimum TPR threshold. The final evaluation shows the impact of training real and synthetic lesion data on metadata-based and image models. In continuation of the work of recent study this implementation improves the the overall melanoma class prediction after combining all the models with better sensitivity and specificity compared to the commercial solutions for skin-lesion detection.

Project 2

Prostate MRI Segmentation using Deep Learning

Checkout - Report, PPT, Github

This project was developed during my Fall 2025 for Image and Video Processing coursework under Prof Yao Wang during NYU MSEE program.

Objectives

Goal was to implement a novel methodology segmenting the Protate158 dataset using MedSAM2 and U-net models.

Results

The work targets anatomy segmentation on the Prostate158 dataset, aiming to reduce the time and inter-reader variability of manual contouring while enabling a fully prompt-less inference path at test time. The paper compares various MedSAM approches and provides a modified architecture for foundation model MedSAM fine-tuned and also using a learned bounding-box predictor (BoxNet), achieving dice more than the recently proposed models on larger dataset.

Page updated

Report abuse