Exploratory multi-cohort, multi-reader study on the clinical utility of a deep learning model for transforming cryosectioned to formalin-fixed, paraffin-embedded (FFPE) images in breast lesion diagnosis.

BACKGROUND

Cryosectioned tissues often exhibit artifacts that compromise pathologists' diagnostic accuracy during intraoperative assessments. These inconsistencies, compounded by variations in frozen section (FS) production across laboratories, highlight the need for improved diagnostic tools. This study aims to develop and validate a deep-learning model that transforms cryosectioned images into formalin-fixed paraffin-embedded (FFPE) images to enhance diagnostic performance in breast lesions.

METHODS

We developed an unpaired image-to-image translation model (AI-FFPE) using the TCGA-BRCA dataset to convert FS images into FFPE-like images. The model employs a modified generative adversarial network (GAN) enhanced with an attention mechanism to correct artifacts and a self-regularization constraint to preserve clinically significant features. For validation, 132 FS whole slide images (WSIs) of breast lesions were collected from three cohorts (SYSUCC, GSPCH, and TCGA). These FS-WSIs were transformed into AI-FFPE-WSIs and independently evaluated by six pathologists for image quality, diagnostic concordance, and confidence in lesion properties and final diagnoses. Diagnostic performance was assessed using a diagnostic score (DS), calculated by multiplying the accuracy index by the confidence level. The dataset included 132 reference diagnoses and 1,584 pathologist reads.

RESULTS

The AI-FFPE group showed a significant improvement in image quality compared to the FS group (p < 0.001). Concordance rates for lesion properties (79.9% vs. 79.9%) and final diagnoses (82.7% vs. 82.6%) were similar between two groups. In concordant cases, the AI-FFPE group demonstrated significantly higher diagnostic confidence than the FS group, with more diagnoses definitively categorized based on lesion properties (54.3% vs. 35.4%, p < 0.001) and final diagnoses (48.3% vs. 33.3%, p < 0.001). Paired t-tests revealed that the diagnostic scores in the AI-FFPE group were significantly higher than in the FS group (overall DS, 13.9 ± 6.6 vs. 12.0 ± 6.6, p < 0.001; DS for lesion property, 6.8 ± 3.8 vs. 5.8 ± 3.7, p < 0.001; DS for final diagnosis, 7.1 ± 3.2 vs. 6.2 ± 3.2, p < 0.001). Logistic regression showed that poorer image quality, atypical ductal hyperplasia/ ductal carcinoma in situ cases, and less experienced pathologists were associated with decreased diagnostic accuracy.

CONCLUSIONS

The AI-FFPE model improved perceived image quality and diagnostic confidence among pathologists in breast lesion evaluations. While diagnostic concordance remained comparable, the enhanced interpretability of AI-FFPE images may support more confident intraoperative decision-making.