Robust Classification of Oral Cancer with Limited Training Data

TL;DR

The paper addresses the challenge of reliable oral cancer screening with scarce labeled data by proposing a hybrid CNN-Bayesian neural network that uses a MobileNet-v1 backbone and a variational dense layer with a spike-and-slab prior to enable uncertainty quantification. By training with variational inference and transfer learning, the approach maintains competitive accuracy on distribution-matched data and superior generalization to distribution-shifted real-world smartphone images, outperforming a traditional CNN when data distributions differ. The study demonstrates calibrated uncertainty: low uncertainty for correct predictions and high uncertainty for misclassifications, enhancing safety and reliability in low-resource healthcare settings. These findings support broader adoption of Bayesian deep learning in healthcare and pave the way for field validation and integration into practical, uncertainty-aware screening tools.

Abstract

Oral cancer ranks among the most prevalent cancers globally, with a particularly high mortality rate in regions lacking adequate healthcare access. Early diagnosis is crucial for reducing mortality; however, challenges persist due to limited oral health programs, inadequate infrastructure, and a shortage of healthcare practitioners. Conventional deep learning models, while promising, often rely on point estimates, leading to overconfidence and reduced reliability. Critically, these models require large datasets to mitigate overfitting and ensure generalizability, an unrealistic demand in settings with limited training data. To address these issues, we propose a hybrid model that combines a convolutional neural network (CNN) with Bayesian deep learning for oral cancer classification using small training sets. This approach employs variational inference to enhance reliability through uncertainty quantification. The model was trained on photographic color images captured by smartphones and evaluated on three distinct test datasets. The proposed method achieved 94% accuracy on a test dataset with a distribution similar to that of the training data, comparable to traditional CNN performance. Notably, for real-world photographic image data, despite limitations and variations differing from the training dataset, the proposed model demonstrated superior generalizability, achieving 88% accuracy on diverse datasets compared to 72.94% for traditional CNNs, even with a smaller dataset. Confidence analysis revealed that the model exhibits low uncertainty (high confidence) for correctly classified samples and high uncertainty (low confidence) for misclassified samples. These results underscore the effectiveness of Bayesian inference in data-scarce environments in enhancing early oral cancer diagnosis by improving model reliability and generalizability.

Figures (5)

• Figure 1: The architecture design for our hybrid Convolutional and Bayesian Neural Network.
• Figure 2: Sample images from the proprietary dataset, cropped Kaggle dataset, and Kaggle dataset.
• Figure 3: Performance analyses plot for our Bayesian model using uncertainty quantification.
• Figure 4: Kernel density estimate (KDE) plot showing the relationship between the prediction distribution and the associated uncertainties for different test datasets. The plot shows how uncertainties vary with the prediction distributions for the datasets under evaluation.
• Figure 5: Uncertainty measurement plots for each test dataset, evaluating the uncertainty for correctly and incorrectly classified samples. The figure displays three plots arranged vertically: the top plot shows the uncertainty measurements for the proprietary test dataset, the middle plot shows the cropped Kaggle test dataset, and the bottom plot shows the Kaggle dataset.