iToBoS 2024 Detection (Kaggle) – Solution Repository

Team

• Solomon Chibuzo Nwafor
• Muhammad Faran Akram

Competition

• iToBoS 2024 – Skin Lesion Detection with 3D-TBP
• Kaggle link: https://www.kaggle.com/competitions/itobos-2024-detection
• NinjaX Result:

Goal

• Detect skin lesions in high resolution skin patch images
• Output bounding boxes per image in the Kaggle submission format

repo content

• Training and inference scripts for a YOLO based detector
• A second stage classifier filter (MedViT) to remove false positives
• Utilities for dataset preparation, split, and evaluation
• Example submissions produced by the pipeline

Approach

1) Detector

• Ultralytics YOLO models (see yolo_train.py)
• Train on YOLO format labels
• Use a 1024 input size at inference (see yolo_test2.py and test_models.py)

2) Data preparation trick

• Blur non skin background to reduce background driven false positives
• Script: preprocess.py
• Output: split_dataset_blurred (configurable inside the script)

3) False positive reduction

• Run YOLO on each test image
• For each predicted box, crop an expanded patch and classify it with MedViT
• Keep the box if
  • YOLO confidence is high (hard keep), or
  • MedViT predicts lesion (keep), else drop
• Script: test_models.py

4) Submission formatting

• prediction_string stores repeated groups of 5 values

  x_min y_min x_max y_max class_id
  

• This repo uses class_id = 0 (single class)
• Scripts producing the CSV
  • yolo_test2.py (YOLO only)
  • test_models.py (YOLO + MedViT filtering)

Repository structure

iToBoS-main/
├── params.yaml
├── preprocess.py
├── split_data.py
├── yolo_train.py
├── yolo_test2.py
├── test_models.py
├── compute_map.py
├── utils.py
├── predictions/
│   ├── filtered_predictions_0.6792.csv
│   └── filtered_predictions_0.6793.csv
├── filtered_predictions1.csv
├── Lesion-Detection-Challange/ (git submodule)
└── MedViT/ (git submodule)

Setup

1) Clone with submodules

git clone --recursive https://github.com/solonso/iToBoS.git
cd iToBoS

2) Install dependencies

• Python 3.9 or newer recommended
• Core packages used in this repo:
  • ultralytics
  • torch
  • torchvision
  • opencv-python
  • numpy
  • pandas
  • pillow
  • tqdm

3) Get the dataset

• Download from the Kaggle competition page
• Prepare folders to match the paths in params.yaml and the scripts
• Labels must be in YOLO format

Configuration

1) params.yaml

• Set dataset paths (train, val, test)
• Set YOLO dataset yaml path (train_yaml)
• Set model list and training parameters if needed

2) Script level constants

These scripts use top of file constants, not CLI flags:

• preprocess.py
  • SOURCE_DIR, DEST_DIR, BLUR_STRENGTH
• split_data.py
  • reads params.yaml
• yolo_train.py
  • reads params.yaml
• yolo_test2.py
  • set model path and test dir inside the file
• test_models.py
  • YOLO weights path, MedViT checkpoint path, output dirs inside the file

Reproduce the pipeline

Step 1: Blur background (optional but used in our runs)

python preprocess.py

Step 2: Create train val split and YOLO directory layout

python split_data.py

Step 3: Train YOLO

python yolo_train.py

Step 4: Generate a submission with YOLO only

python yolo_test2.py

Step 5: Run YOLO + MedViT filter and write final submission

python test_models.py

Outputs

Submissions

• predictions/filtered_predictions_0.6793.csv
• predictions/filtered_predictions_0.6792.csv

If you want quick sanity checks

• Each row is one image_id
• prediction_string is empty for images with no detections
• Each detection uses 5 values (x_min y_min x_max y_max class_id)

Numbers from the example best submission file in this repo

• File: predictions/filtered_predictions_0.6793.csv
• Images in file: 8481
• Average boxes per image: 3.45
• Images with zero boxes: 1944 (22.9 percent)
• Best score label saved with the file name: 0.6793

Credits and references

• Competition and dataset: iToBoS 2024 (Kaggle)
• Official dataset tooling and format converters: iToBoS Lesion Detection Challenge repository
• Dataset description and scale: ISIC archive DOI page for the iToBoS 2024 detection dataset
• Background context on the dataset: arXiv paper describing the iToBoS 3D-TBP derived skin region images