GROUP 5

Colon-cancer-classification

PHASE 2: DL

Hugging Face Hub

• Our best model is loaded at selvaa/vit-colon-cancer-classification

Results

Model	Accuracy	Precision	Recall	F1-Score
ResNet50	76(%)	75(%)	76(%)	75(%)
timm-efficientnetb3	81(%)	81(%)	81(%)	81(%)
YOLO-v8x	80(%)	79(%)	78(%)	78(%)
vit-base-patch16-224	83(%)	83(%)	83(%)	82(%)

Folder structure

Inside phase_2,

• hugging_face_training.py - leverages Hugging Face library to train, validate and pushning model to HF Hub. Current pipeline supports google/vit and facebook/deit variants
• hugging_face_inference.py - loads trained models from HF Hub and does inference on that --> creates predictions csv too
• custom_pipeline/ - contains the training, utils and inference scripts to train torchvision & timm models
• pytorch_lightning/ - uses PyTorch lightning library for easy training, validation, and logging
• report.md - our top3 submission results and their corresponding classification reports

Methodology

For training it using hugging face,

python phase_2/hugging_face_train.py 
#similarly for lightning and custom pipeline

PHASE 1: ML CLASSIFIERS

Best Performing

Features & Results

Feature extraction methods	Computed features
First order statistical features	Mean, Standard deviation, Median, Percentile 25%, Percentile 50%, Percentile 75%
GLCM	Contrast, Dissimilarity, Homogeneity, Energy, ASM, Correlation, Entropy, Variance
Histogram	Histogram of bins=8 (in HSV image space)
Hu invariant moments	h1, h2, h3, h4, h5, h6, h7

Results

Model	Accuracy	Precision	Recall	F1-Score
Voting-based classifier (SVM, XGBoost, LGBM, RF)	80.4(%)	77(%)	76(%)	76(%)
LightGBM	79.2(%)	76(%)	75(%)	76(%)

Folder structure

Inside phase_1,

• previous_best - this folder contains the feature extraction and training pipeline for our previous best approach
• analysis - folder contains scripts for t-SNE and PCA analysis for feature set visualization
• analysis\results - results contains the images that represents the histogram analysis of RGB channels for each type of cancer
• utils.py - feature extraction methods
• train.py - training pipeline for voting based classifier
• submission_pipeline.py - trains and saves the predictions in the submission format

Methodology

For normal training and checking metrics,

python phase_1/train.py

To train the model and generate the predictions CSV file, run

python phase_1/submission_pipeline.py

List of feature extraction methods (we tried):

• GLCM
• Hu invariant moments
• First-order statistical features
• Color histogram (RGB, HSV, LAB)
• LBP
• Gabor filter
• Discrete Fourier Transform (mean, median, std_dev, energy from magnitude spectrum)
• Circular hough transform