± HEALTH Colorectal Cancer (CRC) is a common cause of cancer death and prognostic factors are used to determine manage…+ HEALTH This multicenter study developed and validated an interpretable machine learning model integrating granular n…+ HEALTH Multiple stakeholders need to locate results of registered clinical trials but frequently struggle to find th…+ HEALTH Accurate preoperative glioma grading and molecular subtyping are important for treatment. The vascular microe…+ HEALTH Gastrointestinal endoscopy generates extensive high-resolution video data, posing significant challenges for…+ HEALTH Non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) and invasive encapsulat…+ POLICY Artificial intelligence (AI) has the potential to strengthen real-world evidence (RWE) for regulatory decisio…± HEALTH Echocardiography remains the cornerstone of cardiovascular imaging. However, traditional workflows including…
TruaceTracing the truth around AIFriday, August 28, 2026
Subscribe
Health·G Space·Evidence-backed gain·Published 2026-08-28

Deep Learning-Based Classification of NIFTP and Invasive Encapsulated Follicular Variant of Papillary Thyroid Carcinoma Using Gross Pathology Images

Abstract: Non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) and invasive encapsulated follicular variant of papillary thyroid carcinoma (IEFVPTC) are diagnostically challenging thyroid neoplasms with overlapping clinical and molecular characteristics. Although artificial intelligence has shown promise for diagnostic support in radiology and histopathology, its application to gross pathology remains unexplored. This study analyzed gross pathology photographs from 87 patients (43 with NIF…

TRV-2026-0915Peer-reviewedPermanent record — cite & verify
Deep Learning-Based Classification of NIFTP and Invasive Encapsulated Follicular Variant of Papillary Thyroid Carcinoma Using Gross Pathology Images

"Asbestosis- Asbestos bodies Case 194" by Pulmonary Pathology is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.

The quick read

On August 26, 2026, a peer-reviewed study reported a proof-of-concept deep learning approach to distinguish two diagnostically challenging thyroid neoplasms, NIFTP and IEFVPTC, using gross pathology photographs from 87 patients. Using frozen pretrained CNN backbones and traditional classifiers with nested cross-validation, the prespecified EfficientNet plus Random Forest model achieved a pooled AUC of 0.788.

The work matters because it extends AI diagnostic support beyond radiology and histopathology into gross pathology, suggesting macroscopic morphology encodes subtle class-separating signals. Uncertainty remains due to the small single-cohort sample, modest sensitivity of 0.588, and lack of significant differences among models after correction, indicating need for larger external validation before clinical use.

Main points
  • Study included 87 patients: 43 with NIFTP and 44 with IEFVPTC using gross pathology photographs.
  • Three frozen pretrained CNN backbones (EfficientNet, ResNet, ConvNeXt) generated embeddings classified by Random Forest, SVM, and Gradient Boosting with patient-level nested stratified five-fold cross-validation.
  • Primary model sensitivity 0.588, specificity 0.944, accuracy 0.771, F1 0.714; exploratory combinations had pooled AUCs 0.500 to 0.755 with no significant pairwise differences after Benjamini-Hochberg correction.
  • Grad-CAM showed backbones predominantly highlighted lesional tissue rather than background regions, though activation patterns differed across architectures.
Gain

A prespecified EfficientNet plus Random Forest model classified gross pathology photographs to distinguish NIFTP from IEFVPTC with pooled AUC 0.788, sensitivity 0.588 and specificity 0.944 in 87 patients.

The rundown

Researchers used frozen pretrained EfficientNet, ResNet, and ConvNeXt backbones to extract features from gross pathology photographs of 87 thyroid cases, then trained Random Forest, SVM, and Gradient Boosting classifiers with patient-level nested stratified five-fold cross-validation and inner hyperparameter tuning.

The prespecified EfficientNet plus Random Forest model yielded pooled AUC 0.788 with high specificity but lower sensitivity, while eight exploratory combinations ranged from 0.500 to 0.755 and Grad-CAM indicated attention focused on lesional tissue rather than background.

Sources

Reader signal

How should this claim be treated?

The debate