Search

Simple Summary: Lip and oral cavity cancer caused over 177,000 deaths globally in 2020, but patient survival increases with earlier diagnosis. One barrier to early diagnosis is the invasive nature of biopsies needed for diagnosis. Automated diagnosis systems have the potential to perform non-invasive diagnosis by pairing novel imaging data with deep learning models. Given the variability between patients, access to a sufficiently large training database from human subjects limits deep learning applications. We propose a model that maps non-invasive images of oral tissue to a diagnosis by encouraging the model to group normal samples close together (reducing variability between patients). Our model improves non-invasive oral cancer diagnosis through a robust training process that only requires a small amount of data. This work shows how we can address small data challenges through model architecture and training, rather than through the collection of larger databases or manual corrections and normalizations. Background: Multispectral autofluorescence lifetime imaging systems have recently been developed to quickly and non-invasively assess tissue properties for applications in oral cancer diagnosis. As a non-traditional imaging modality, the autofluorescence signal collected from the system cannot be directly visually assessed by a clinician and a model is needed to generate a diagnosis for each image. However, training a deep learning model from scratch on small multispectral autofluorescence datasets can fail due to inter-patient variability, poor initialization, and overfitting. Methods: We propose a contrastive-based pre-training approach that teaches the network to perform patient normalization without requiring a direct comparison to a reference sample. We then use the contrastive pre-trained encoder as a favorable initialization for classification. To train the classifiers, we efficiently use available data and reduce overfitting through a multitask framework with margin delineation and cancer diagnosis tasks. We evaluate the model over 67 patients using 10-fold cross-validation and evaluate significance using paired, one-tailed t-tests. Results: The proposed approach achieves a sensitivity of 82.08% and specificity of 75.92% on the cancer diagnosis task with a sensitivity of 91.83% and specificity of 79.31% for margin delineation as an auxiliary task. In comparison to existing approaches, our method significantly outperforms a support vector machine (SVM) implemented with either sequential feature selection (SFS) (p = 0.0261) or L1 loss (p = 0.0452) when considering the average of sensitivity and specificity. Specifically, the proposed approach increases performance by 2.75% compared to the L1 model and 4.87% compared to the SFS model. In addition, there is a significant increase in specificity of 8.34% compared to the baseline autoencoder model (p = 0.0070). Conclusions: Our method effectively trains deep learning models for small data applications when existing, large pre-trained models are not suitable for fine-tuning. While we designed the network for a specific imaging modality, we report the development process so that the insights gained can be applied to address similar challenges in other non-traditional imaging modalities. A key contribution of this paper is a neural network framework for multi-spectral fluorescence lifetime-based tissue discrimination that performs patient normalization without requiring a reference (healthy) sample from each patient at test time. [ABSTRACT FROM AUTHOR]