Predicting oral cancer risk using WSI and DL

TITLE

Deep learning-based pathology image analysis predicts cancer progression risk in patients with oral leukoplakia

AUTHORS

Zhang X; Gleber-Netto FO; Wang S; Martins-Chaves RR; Gomez RS; Vigneswaran N; et al.

SUMMARY

This study developed a convolutional neural network (CNN) model, named OMRS (Oral Mucosa Risk Stratification), to predict the risk of malignant transformation in oral leukoplakia (OL) patients. The model was trained on H&E-stained images of non-dysplastic oral mucosa and oral squamous cell carcinoma (OSCC) to identify morphological features associated with cancer progression. When applied to a cohort of 62 OL patients, the OMRS model successfully stratified patients into high- and low-risk groups, with high-risk patients showing a significantly higher probability of developing oral cancer (HR = 3.98 to 4.52) compared to low-risk patients. The model outperformed the traditional WHO grading system in predicting progression.

EXTRACTED DATA:

Year: 2023

Modality: Histopathology (H&E) – Whole Slide Imaging (WSI)

Dataset: 38 WSI for model development (13 nondysplastic oral mucosa + 25 OSCC from TCGA) + 62 WSI of oral leukoplakia for prognostic evaluation

Dataset Split: Patches: - Training (2419 OSCC / 2981 Oral mucosa) - Validation (451 OSCC / 709 Oral mucosa) - Testing (513 OSCC / 739 Oral mucosa)

Architecture: Modified Inception v3 (CNN)

Metrics: Patch-level: Accuracy (95.4%), ROC-AUC (0.992–0.999); Patient-level: Hazard Ratio (HR), p-value, 5-year progression probability

AP Qty: 1

Supervisor: No information

Expertise: OMFP

Tool/System: Unavailable (Manual evaluation to annotate areas mentioned, but specific software not named)

ML Task: Patch-level classification with slide-level aggregation for cancer progression risk prediction

Objective: To predict malignant transformation risk and time-to-progression in oral leukoplakia

CLINICAL RELEVANCE

The OMRS model addresses the limitations of the current WHO histopathological grading system, which is subjective and suffers from inter-observer variability. By providing an automated, objective risk score, the model identified high-risk patients who were approximately 4 times more likely to develop oral cancer than low-risk patients, even after adjusting for clinical variables. This tool has the potential to guide clinical management by allowing for tailored surveillance intervals, safeguarding high-risk patients while reducing unnecessary interventions for low-risk patients.