26th Australian Institute of Physics Congress and the Australian and New Zealand Conference on Optics and Photonics

High-Sensitivity Deep Learning-Enhanced Raman Spectral Classification for Prostate Cancer Detection

Not scheduled

20m

Belinda Hutchinson Building (The University of Sydney )

Contributed Oral ANZOS | Photonics and Optics (ANZCOP)

Speaker

Hanlin Li (The University of Auckland)

Description

Prostate cancer (PCa) is the most diagnosed cancer among males and the second leading cause of cancer death in males in New Zealand. Current diagnosis still relies on biopsy-based tissue sampling and histopathological assessment. During prostate cancer surgery, no tools can assess prostate tissue in real time, thus leading to a high positive surgical margin rate. Raman spectroscopy (RS), combined with classification models, is well positioned to address these needs because it can capture biochemical information from tissue rapidly. Previous Raman-based prostate cancer studies have mainly used conventional classification models, which have shown promising performance but may be limited in detecting subtle, multi-scale spectral differences. CNN-based models can learn diagnostic spectral features more directly, but their use in prostate tissue Raman classification remains limited.

This work developed a Multi-scale Attention-Residual Spectral Convolutional Neural Network (MARS-CNN) for cancer/benign classification. The model was trained and tested using 6,142 ex vivo Raman spectra acquired from prostate biopsy tissue from 152 patients. To capture diagnostic information that may appear as small peak changes, broader band patterns, or relative intensity differences between Raman bands, MARS-CNN uses 5-, 7-, and 10-point convolutional kernels, residual pathways, and channel attention to examine multi-scale spectral features, retain weak Raman signals, and focus on cancer-related patterns.

MARS-CNN achieved an AUC of 0.911, with 91.4% sensitivity, 80.5% specificity, and a negative predictive value of 96.0%, representing the best performance reported on prostate cancer tissue compared with the other evaluated models, including a basic CNN. These results indicate high cancer detection while maintaining reliable benign classification, reducing both missed malignancy and false-positive assessment. These findings support Raman-based assessment as a practical method to guide more targeted prostate tissue evaluation, with potential for future real-time clinical use.