Source-linked AI summary
Learning to Measure and Diagnose: A Clinical-Dictionary Radiomic Bottleneck Model for Thyroid Nodule Classification
Guibin Zhang, Weier Luo, Lei Cao
TL;DR
Thyroid ultrasound classifiers can predict malignancy without showing evidence linked to familiar clinical descriptions. This model predicts clinical-dictionary radiomic measurements from images and classifies from selected estimates, with competitive results reported on ThyroidXL and TN3K.
Problem
Thyroid nodule classification predictions may not reveal the ultrasound findings supporting them, while using radiomic measurements directly for classification requires segmentation at inference.
Source: §1 · pp. 1–2Method
The model predicts radiomic measurements from ultrasound, balances supervision across feature families, and jointly optimizes estimation and classification under fidelity guidance, tracing feature contributions by dictionary category.
Source: §1 · p. 2, §2.1 · p. 2, §2.4 · p. 4Results
The method achieves competitive classification and the highest mean AUC and BACC among interpretable methods on both ThyroidXL and TN3K.
Source: §3.2 · p. 6, §4 · p. 7Takeaways & Limitations
Individual predictions have an inspectable quantitative basis through traceable feature contributions associated with ultrasound descriptions.
Source: §4 · p. 7
Abstract
from arXiv · showhide
Thyroid ultrasound supports the assessment of nodule malignancy, but classification predictions often lack evidence linked to familiar clinical descriptions. Radiomics provides explicitly defined measurements that can support such explanations through a clinical dictionary. However, using these measurements as interpretable evidence requires reliable estimation from ultrasound images and preservation of measurement fidelity when optimizing classification performance. We propose a clinical-dictionary radiomic bottleneck model for thyroid nodule classification. Within this model, Family-Aware Radiomic Learning (FARL) estimates measurements with balanced supervision across feature families. Fidelity-Guided Bottleneck Modeling (FGBM) constructs a fixed bottleneck, followed by joint optimization of measurement estimation and classification under fidelity guidance. Decision evidence decomposition then traces feature contributions through a linear classifier and organizes them by clinical dictionary categories. Experiments on ThyroidXL and TN3K show competitive classification performance, while case studies illustrate feature contributions for benign and malignant nodules. These findings indicate interpretable classification with inspectable quantitative evidence informed by clinical knowledge. Code is available online.
1 Introduction
The introduction argues that thyroid nodule predictions need inspectable evidence linked to familiar ultrasound descriptions. It proposes predicting clinically associated radiomic measurements from ultrasound images and using them for classification, while balancing measurement reliability and diagnostic performance.
1 Introduction
- Classification predictions alone do not reveal which ultrasound findings support a prediction, motivating evidence that is inspectable and linked to familiar descriptions.Clinicians use ultrasound findings to assess which nodules warrant biopsy, follow-up, or no further evaluation.Source: §1 · p. 1
- Existing concept and prototype explanations may not align with established ultrasound descriptors (Oikarinen et al., 2023).The cited approaches map images to interpretable concepts or relate image regions to prototypes and part-based concepts.Source: §1 · p. 1
- A clinical dictionary associates radiomic measurements with five ultrasound descriptor categories: composition, echogenicity, shape, margin, and echogenic foci.The measurements describe lesion geometry, intensity, and texture, and their explicit definitions allow predicted values to be checked against measured references.Source: §1 · pp. 1–2
- Radiomic feature families may differ in learning difficulty, while classification performance and fidelity to measured references are not necessarily aligned.Lower overall estimation error may not improve diagnostic discrimination, and classification optimization alone may compromise measurement fidelity.Source: §1 · p. 2
- The proposed model predicts radiomic measurements from ultrasound and classifies from them, combining FARL, fidelity-guided FGBM, and category-organized decision evidence.The authors report competitive classification on ThyroidXL and TN3K, with case-level analyses of benign and malignant nodules.Source: §1 · p. 2
2 Method
The model predicts radiomic measurements from ultrasound images and uses selected estimates in a fixed diagnostic bottleneck for classification. It combines balanced family supervision, fidelity-guided joint optimization, and category-organized decision evidence.
2.1 Model Overview
- The model predicts radiomic measurements from ultrasound images, selects a fixed dictionary bottleneck, jointly optimizes estimation and classification, and traces linear-classifier contributions by clinical category.FARL estimates the full radiomic feature vector; FGBM uses selected estimates for classification.Source: §2.1 · p. 2
2.2 Family-Aware Radiomic Learning
- FARL averages SmoothL1 errors within geometry, intensity, and texture families, then weights family losses equally to prevent larger families from dominating supervision.Its targets are offline radiomic measurements, and its per-image predictions are inverse-transformed and averaged into case-level estimates.Source: §2.2 · pp. 2–3
2.3 Fidelity-Guided Bottleneck Modeling
- FGBM selects clinically associated features by reliability while limiting redundancy, then keeps their combined bottleneck fixed during subsequent optimization.It jointly optimizes feature estimation and classification with standardized selected estimates passed to a linear classifier; fidelity is anchored to frozen FARL predictions.Source: §2.3 · p. 4
2.4 Decision Evidence Decomposition
- The decision score is decomposed into selected-feature contributions and organized by associated dictionary categories, with multi-category feature contributions distributed equally.The category contributions exactly reconstruct the decision logit; they organize diagnostic evidence rather than represent independent predictions of ultrasound findings.Source: §2.4 · p. 4
- RCAF and Seg-first radiomics use predicted masks, whereas TU1.0 and Oracle radiomics require ground-truth masks at inference.Source: §2.4 · p. 5
3 Experiments and Results
The experiments evaluate the model on ThyroidXL and TN3K, compare it with other classification approaches, test FARL and FGBM design choices, and illustrate feature contributions for representative cases.
3.1 Experimental Setup
- Evaluation used ThyroidXL and TN3K with fixed train, validation, and test splits, patient-level evaluation for ThyroidXL, and image-level evaluation for TN3K.The datasets contained 11,554 and 3,493 image–mask pairs, respectively.Source: §3.1 · p. 5
3.2 Comparison Experiment
- Our method achieved mean AUC and BACC of 94.77% and 86.73% on ThyroidXL, and 79.49% and 72.23% on TN3K; it had the highest mean AUC and BACC among interpretable methods on both datasets.It outperformed all three radiomics references on TN3K, while TU1.0 and Oracle radiomics performed better on ThyroidXL but required ground-truth masks at inference.Source: §3.2 · pp. 5–6
3.3 Ablation Study
- Removing joint optimization reduced AUC by 2.57 percentage points on ThyroidXL and 6.25 percentage points on TN3K; omitting family balancing slightly reduced mean AUC on both datasets.Without the fidelity loss, mean Spearman correlation decreased on both datasets alongside lower AUC.Source: Table 2 · p. 6, §3.3 · p. 6
3.4 Case Study
- Figure 2’s benign and malignant cases had different dominant category contributions: echogenicity and margin for benign, versus shape and composition for malignant nodules.GLCM Id reflected relatively uniform benign texture; Shape2D Elongation and GLSZM SmallAreaEmphasis reflected malignant nodule elongation and small homogeneous zones.Source: §3.4 · p. 6
4 Conclusion
The paper presents a clinical-dictionary radiomic bottleneck model for thyroid ultrasound classification, with performance balanced against measurement fidelity. Its quantitative feature contributions are inspectable and associated with ultrasound descriptions.
- The model balances classification performance with measurement fidelity and makes the quantitative basis of individual predictions inspectable through traceable feature contributions associated with ultrasound descriptions.The authors identify extending the model to more medical scenarios using task-specific quantitative measurements and clinical knowledge as future work.Source: §4 · pp. 6–7