vatche-t / polyp-segmentation

Polyp Segmentation Project Report

Problem Statement

This project aims to develop a machine learning model for segmenting gastrointestinal polyps in medical images, addressing the question of whether we can automate the identification and delineation of polyps to assist healthcare professionals in early diagnosis and treatment planning. The project seeks to add societal value by reducing manual segmentation efforts, thus expediting diagnostics and improving patient outcomes.

Background

Segmentation of gastrointestinal polyps is crucial for early detection of conditions potentially leading to colorectal cancer. Traditional manual inspection methods are time-consuming and error-prone. Utilizing machine learning and computer vision can automate this process, enhancing speed and accuracy.

Dataset

• Source and Structure: The project uses the Kvasir-SEG dataset, featuring gastrointestinal polyp images and masks in JPEG format, along with bounding box data in JSON format.
• Collection: Collected during colonoscopy procedures by medical professionals, aimed at aiding computer-aided diagnosis research.
• Schema:
  • Images in JPEG, representing gastrointestinal polyps.
  • Segmentation Masks in JPEG, serving as ground truth.
  • Bounding Boxes in JSON, detailing polyp locations.

Cleaning and Preprocessing

• Conducted Exploratory Data Analysis (EDA) for understanding polyp sizes, locations, and color distributions.
• Preprocessing involved rescaling pixel values and resizing images/masks to 256x256 pixels.
• Applied data augmentation techniques like flips, rotation, zooming, and brightness adjustments to enhance model generalizability.

Insights, Modeling, and Results

• EDA Insights: Identified polyp distribution across images and quantified sizes, revealing a tendency for central placement and varied sizes.
• Model: Employed U-Net architecture for its efficacy in biomedical image segmentation.
• Evaluation: Used accuracy, precision, recall, F1-score, IoU, and Dice Coefficient, achieving 95% accuracy, 91% precision for polyps, and significant overlap between predicted and ground-truth masks.

Conclusions

The project validates the hypothesis that advanced machine learning techniques can produce a high-accuracy model for polyp segmentation, supporting early diagnosis and treatment planning. Despite its efficacy, optimization opportunities exist, especially in reducing false positives and improving spatial accuracy.

Practical Applications and Future Directions

• Practical Applications: Integration into medical imaging systems for real-time segmentation, aiding in early and accurate diagnosis.
• Future Directions: Focus on model optimization and extension to other medical imaging and segmentation tasks.

Last updated on: 2024-02-11

Last updated on: 2024-02-14

Last updated on: 2024-02-15

Last updated on: 2024-02-18

Last updated on: 2024-02-19

Last updated on: 2024-02-20

Last updated on: 2024-02-20

Last updated on: 2024-02-28

Last updated on: 2024-03-04

Last updated on: 2024-03-05

Last updated on: 2024-03-18

Last updated on: 2024-03-22

Last updated on: 2024-03-28

Last updated on: 2024-03-30

Last updated on: 2024-04-08

Last updated on: 2024-04-16

Last updated on: 2024-04-19

Last updated on: 2024-04-23

Last updated on: 2024-04-29

Last updated on: 2024-05-01

Last updated on: 2024-05-11