histopathologic-cancer-detection

Histopathologic Cancer Detection

Prerequisites: Ensure you have installed mlebench following instructions under the MLE Bench repo.

Overview

Histopathologic cancer detection is a critical task in medical image analysis that involves identifying cancerous cells in microscopic images of tissue samples. Early and accurate detection of cancer cells can significantly improve patient outcomes and treatment planning. This task focuses on developing machine learning models to automate the detection of cancer cells in histopathology images, which can help pathologists work more efficiently and reduce human error.

In this task, you will develop a model to identify cancer cells in histopathology images. The challenge involves analyzing microscopic images of tissue samples and determining whether they contain cancerous cells. This is a binary classification task where each image needs to be classified as either containing cancer cells (positive) or not (negative).

Download Dataset

mlebench prepare -c histopathologic-cancer-detection

Run Curie

1. Update the configuration: Open curie/configs/mle-histopathologic-cancer.json and verify the paths to the dataset and starter code.

2. Execute Curie:

cd Curie/
python3 -m curie.main -f benchmark/mle_bench/histopathologic-cancer-detection/histopathologic-cancer-detection.txt --task_config curie/configs/mle_config.json --dataset_dir /home/amberljc/.cache/mle-bench/data/histopathologic-cancer-detection/prepared/public

3. Change --dataset_dir to the absolute path to your dataset.

Curie Results

After asking Curie to solve this question, the following output files are generated:

• Auto-generated Report
• Experiment results: All detailed results for all conducted experiments
• Curie logs: Execution log file
• Curie workspace: Generated code, complete script to reproduce and raw results (we didn't upload the model checkpoint).

Curie Performance Summary

Model	AUC-ROC	Training Time (s)	Inference Time (ms/sample)	Model Size (MB)
ResNet50	0.9727	34.37	0.61	89.89
DenseNet121	0.9778	57.02	0.65	26.85
EfficientNetB0	0.9840	33.09	0.62	15.45
SEResNeXt50	0.9807	48.56	0.62	97.58
Custom w/Attention	0.9840	60.17	0.61	95.88

Dataset

The dataset consists of histopathology images from the PatchCamelyon (PCam) benchmark dataset, which is derived from the Camelyon16 challenge. The images are 96x96 pixel patches extracted from larger whole-slide images of lymph node sections.

• Images: Available in PNG format (96x96 pixels)
• Metadata:
  • id: Unique identifier for each image
  • label: Binary classification label (0 = negative, 1 = positive for cancer)

Dataset Characteristics

• Image Size: 96x96 pixels
• Color Channels: RGB
• Total Images: ~220,000 training images
• Class Distribution: Approximately balanced between positive and negative cases
• Image Source: Lymph node sections from whole-slide images

Dataset Challenges

• Small Image Size: The 96x96 pixel patches require efficient feature extraction
• Complex Patterns: Cancer cells can appear in various forms and patterns
• Image Quality: Variations in staining and tissue preparation can affect image appearance
• Computational Efficiency: Large dataset size requires efficient training approaches