Contrastive Clustering-Based Patient Normalization to Improve Automated In Vivo Oral Cancer Diagnosis from...

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Optical Diagnostics

Background:

• Multispectral autofluorescence lifetime imaging (MAFI) offers non-invasive tissue assessment for oral cancer detection.
• Clinical interpretation of MAFI data is challenging due to its non-traditional nature, necessitating diagnostic models.
• Training deep learning models on small MAFI datasets is hindered by patient variability and overfitting.

Purpose of the Study:

• To develop a robust deep learning framework for oral cancer diagnosis using MAFI data.
• To address challenges of small datasets and inter-patient variability in training diagnostic models.
• To enable patient-specific normalization without requiring a reference sample.

Main Methods:

• A contrastive-based pre-training strategy was employed for patient normalization.
• A multitask learning framework incorporated margin delineation and cancer diagnosis.
• The pre-trained encoder was used for classification, evaluated via 10-fold cross-validation on 67 patients.

Main Results:

• The proposed method achieved 82.08% sensitivity and 75.92% specificity for cancer diagnosis.
• An auxiliary margin delineation task yielded 91.83% sensitivity and 79.31% specificity.
• Significant performance improvements were observed compared to SVM and autoencoder baseline models.

Conclusions:

• The developed deep learning approach effectively trains models for small data applications in non-traditional imaging.
• The framework enables patient normalization without reference samples, a key advancement for MAFI-based diagnostics.
• Insights from this study are applicable to similar challenges in other novel imaging modalities.