🤖 Generated with AI assistance · Does not replace individualized medical advice.

AI distinguishes NIFTP from papillary thyroid carcinoma with macroscopic images: a modest but promising first step

Keyfacts

Context: why does this distinction matter?

The NIFTP (noninvasive follicular thyroid neoplasm with papillary-like nuclear features) was redefined in 2016 as a benign entity, separating it from the encapsulated follicular variant of papillary thyroid carcinoma (IEFVPTC) PMID: 27466257. This reclassification has a direct impact on clinical management: NIFTP is treated with total thyroidectomy or lobectomy without the need for prophylactic lymphadenectomy or radioactive iodine therapy, whereas IEFVPTC, being invasive, requires a more aggressive oncological approach.

However, the distinction between both entities is notoriously difficult, even for expert pathologists. The papillary nuclear features overlap, and capsular or vascular invasion can be subtle, requiring exhaustive sampling and meticulous microscopic analysis. In this scenario, any tool that can provide additional diagnostic information, especially in the initial phase of macroscopic evaluation, is of great interest.

So far, artificial intelligence has been successfully applied in radiology and digital histopathology, but its use in macroscopic pathology (photographs of the surgical specimen) remained unexplored. This study by Tsai et al. [DOI: 10.1007/s10278-026-02236-z] addresses that gap, hypothesizing that macroscopic morphology encodes subtle but discriminative signals between NIFTP and IEFVPTC.

Study details

The authors analyzed macroscopic pathology photographs from 87 patients (43 with NIFTP and 44 with IEFVPTC). They used three pre-trained convolutional neural network architectures (EfficientNet, ResNet, and ConvNeXt) to generate feature embeddings, which were then classified using three supervised learning algorithms: Random Forest, support vector machine (SVM), and Gradient Boosting. Validation was performed using five-fold nested stratified cross-validation with internal hyperparameter tuning, ensuring unbiased out-of-fold predictions.

The model prescribed as the primary model was EfficientNet combined with Random Forest, which achieved an AUC of 0.788 (95% CI: 0.614–0.928) [DOI: 10.1007/s10278-026-02236-z]. The remaining eight exploratory models (combinations of the three backbones with the three classifiers) showed AUCs between 0.500 and 0.755, with no statistically significant differences after Benjamini-Hochberg correction [DOI: 10.1007/s10278-026-02236-z].

The primary model presented a sensitivity of 0.588, a specificity of 0.944, a precision of 0.771, and an F1 score of 0.714 [DOI: 10.1007/s10278-026-02236-z]. Grad-CAM analysis revealed that all architectures predominantly highlighted the lesion tissue, not the background, although the activation patterns varied according to the backbone.

Model Performance