Skin Cancer Detection with 3D-TBP

Image Classification

Multimodal

Ensembling

Identified cancers among skin lesions cropped from 3D total body photographs along with metadata

The International Skin Imaging Collaboration (ISIC) is an international academia and industry partnership designed to reduce skin cancer morbidity and mortality through the development and use of digital skin imaging applications. Beginning in 2016, ISIC has sponsored Grand Challenges for the computer science community in association with leading computer vision conferences. Over the years, these challenges have grown in scale, complexity, and participation.

Problem Statement

The task is to develop image-based algorithms to identify histologically confirmed skin cancer cases with single-lesion crops from 3D total body photos (TBP). The image quality resembles close-up smartphone photos, which are regularly submitted for telehealth purposes. The binary classification algorithm could be used in settings without access to specialized care and improve triage for early skin cancer detection.

Data

The dataset consists of diagnostically labelled images with additional metadata. The images are JPEGs. The associated .csv file contains a binary diagnostic label (target), potential input variables (e.g. age_approx, sex, anatom_site_general, etc.), and additional attributes (e.g. image source and precise diagnosis). In the challenge, one is to differentiate benign from malignant cases. For each image (isic_id) one is to assign the probability (target) ranging [0, 1] that the case is malignant. To download the data files, please run the following shell command:

kaggle competitions download -c isic-2024-challenge

Evaluation

Submissions are evaluated on partial area under the ROC curve (pAUC) above 80% true positive rate (TPR) for binary classification of malignant examples. The receiver operating characteristic (ROC) curve illustrates the diagnostic ability of a given binary classifier system as its discrimination threshold is varied. However, there are regions in the ROC space where the values of TPR are unacceptable in clinical practice. Systems that aid in diagnosing cancers are required to be highly-sensitive, so this metric focuses on the area under the ROC curve AND above 80% TRP. Hence, scores range from [0.0, 0.2].

Approach

I adopted a structured and multi-layered methodology to achieve significant performance improvements.

Data Pre-processing

The task involved tackling an imbalanced binary classification problem, where the training dataset had a class ratio of 1:1000 (positive to negative). To address this, I first undersampled the negative class and then oversampled the positive class, achieving a balanced ratio of 2:3. Additionally, I incorporated penalties in the loss function to account for the adjusted class. To enhance the dataset, I applied various image augmentation techniques and relied extensively on the fastai library for efficient data preprocessing and augmentation workflows.

Fine-tuning the Base Model

I fine-tuned a pre-trained ResNet-18 model on the single lesion crops. I utilized StratifiedGroupKFold to ensure robust evaluation and also to generate Out-Of-Fold (OOF) predictions.

Building an Ensemble System

Leveraging the OOF predictions from the ResNet-18 image model, I integrated additional predictive layers by stacking three powerful gradient boosting models: XGBoost, CatBoost, and LightGBM. Each of these models independently processed the image features, and their outputs were combined using a soft voting ensemble. This ensemble strategy balanced the strengths of individual models, yielding a more generalized and accurate prediction system.

Result

The ensemble approach resulted in a significant improvement in the partial AUC-ROC score. The metric increased to 0.167, outperforming the standalone image model’s score of 0.144, thus showcasing the enhanced predictive capability of the multi-model ensemble.