Attention-based image segmentation and classification model for the preoperative risk stratification of thyroid nodules

Karishma Jassal; Bruno Di Muzio; Melissa Edwards; Wendy Brown; Jonathan Serpell; Afsaneh Koohestani; James C Lee

doi:10.1002/wjs.12464

Attention-based image segmentation and classification model for the preoperative risk stratification of thyroid nodules

World J Surg. 2024 Dec 29. doi: 10.1002/wjs.12464. Online ahead of print.

Authors

Karishma Jassal^{1

2}, Bruno Di Muzio³, Melissa Edwards⁴, Wendy Brown², Jonathan Serpell^{1

2}, Afsaneh Koohestani^{1

2}, James C Lee^{1

2}

Affiliations

¹ Monash University Endocrine Surgery Unit, Department of General Surgery, Alfred Hospital, Melbourne, Victoria, Australia.
² Department of Surgery, Central Clinical School, Monash University, Melbourne, Victoria, Australia.
³ Department of Radiology, Alfred Hospital, Melbourne, Victoria, Australia.
⁴ Department of Surgery, Peter MacCallum Cancer Centre, Melbourne, Victoria, Australia.

PMID: 39734138
DOI: 10.1002/wjs.12464

Abstract

Background: Despite widespread use of standardized classification systems, risk stratification of thyroid nodules is nuanced and often requires diagnostic surgery. Genomic sequencing is available for this dilemma however, costs and access restricts global applicability. Artificial intelligence (AI) has the potential to overcome this issue nevertheless, the need for black-box interpretability is pertinent. We aimed to create an ultrasonographic segmentation and classification model that offers explainability and risk accountability.

Methodology: Four hundred and fourteen ultrasonography images were collected from 105 patients undergoing thyroidectomy, divided into training and testing groups. Classification ground truth used is exclusively surgical histopathology. Relevant nodules were manually annotated by a dedicated study radiologist and surgeon. Three AI architectures with and without block attention modules were trained to identify the relevant nodule and the best performing was selected for the subsequent task in classifying identified nodules into benign or malignant. Gradient-Weighted Class Activation Map is used to provide saliency mapping for visual interpretability.

Findings: Superior performance was recorded by the block attention model which stratified thyroid nodules into benign versus malignant with an accuracy of 93% versus 90%, F-score 90% versus 89%, sensitivity 93% versus 91% and specificity 92% versus 91% on a training dataset versus a testing dataset respectively.

Gradcam: Visual interpretability maps demonstrate salient areas for a benign nodule diagnosis overlaps spongiform areas and malignant diagnosis salient areas overlap solid components of a partially cystic-solid nodule and microcalcifications within nodules. These findings are consistent with established diagnostic criteria for benign and malignant nodules.

Conclusion: We developed an image segmentation and classification model for the risk stratification of thyroid nodules benchmarking surgical histopathology as ground truth and providing visual interpretability.

Keywords: endocrine.