A Controlled Audit of Architectural Complexity in Uncertainty-Aware Multi-Organ Ultrasound Classification

Yang Song; Pengbo Sun; Shichang Feng; Ye Zhu; Xin Xu; Ziran Wang

cs.CVAug 28, 2026

A Controlled Audit of Architectural Complexity in Uncertainty-Aware Multi-Organ Ultrasound Classification

Authors: Yang Song, Pengbo Sun, Shichang Feng, Ye Zhu, Xin Xu, Ziran Wang

Organizations: Department of Mechanical Engineering, The Hong Kong Polytechnic University, Kowloon, Hong Kong · School of Electronic and Information Engineering, Beihang University, Beijing, 100191, China · Hangzhou International Innovation Institute, Beihang University, Hangzhou, 311115, China · College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Nanjing, China · Independent Researcher, Jinan, Shandong, China · Key Laboratory of High Efficiency and Clean Mechanical Manufacture, Ministry of Education, School of Mechanical Engineering, Shandong University, 250000, Jinan, China

Abstract

Multi-organ ultrasound classifiers increasingly combine attention, mixture-of-experts routing, uncertainty gating, and evidential deep learning (EDL) objectives to address heterogeneous anatomy and acquisition. Yet a plausible design rationale does not by itself establish that an added component improves the trained system. We contribute a controlled complexity-audit framework, applied to the deployment decision between the maximal evidential candidate Full-EDL and simpler alternatives. Six candidates were evaluated on the primary dataset and three in an internal replication, using ten matched seeds, frozen image-level partitions, capacity- and optimisation-aware comparisons, symmetric temperature scaling, paired decision rules, and a separate out-of-distribution (OOD) veto. Retaining Full-EDL did not establish a reliable macro-F1 gain on either dataset, while the simplified alternatives remained inconclusive under the non-inferiority margin. Simple cross-entropy with temperature scaling (Simple-CE+TS) met the calibrated negative log-likelihood criterion on both datasets and showed favourable selective-risk ordering. The raw calibration advantage of evidential training disappeared after temperature scaling and did not recur on the second dataset. The gate had negligible observable influence at the audited checkpoints, and deleting the Full-only chain revealed no stable task or calibrated-loss benefit. Simple-CE nevertheless triggered the OOD veto against the fetal probe but not the lung probe, precluding an unconditional OOD-safety claim. We therefore selected Simple-CE+TS for the evaluated in-distribution objective while retaining Full-EDL as the maximal reference. Components should earn retention through functional and retraining-based evidence, and calibration and distribution-shift reliability should be evaluated separately.

Figures & tables

Figure 1: Controlled candidate hierarchy and engineering-selection logic. (a) All candidates share an ImageNet-pretrained MobileNetV3-Small feature extractor. (b) The evaluated design space separates structural complexity (Simple, Expert-only, and Full) from the training objective (CE or EDL). Expert-only-CE was not evaluated, and Slim-CE is a secondary lower-capacity operating point. (c) The audit combines matched optimisation, capacity-aware comparison, ten paired seeds, symmetric calibration, fixed-checkpoint gate counterfactuals, physical module deletion, and endpoint-specific decisions. Simple-CE+TS is selected for the in-distribution objective, whereas Full-EDL is retained as the maximal reference and OOD safety is evaluated separately.

Candidate	Width	NALA	Probe	Router+4 experts	Shared+gate	Objective	Params (M)	MACs (M)	Evaluated on
Slim-CE	256	No	No	No	No	CE	1.229	69.4	D2
Simple-CE	1280	No	No	No	No	CE	2.411	127.2	D1, D2
Simple-EDL	1280	No	No	No	No	EDL	2.411	127.2	D1, D2
Expert-only-EDL	256	No	No	Yes	No	EDL	2.121	112.7	D2
Full-CE	256	Yes	Yes	Yes	Yes	CE	2.438	129.9	D2
Full-EDL	256	Yes	Yes	Yes	Yes	EDL	2.438	129.9	D1, D2

Table 1: Candidate design matrix. All candidates use the same MobileNetV3-Small feature extractor. Temperature scaling is applied symmetrically after training and is not an architectural component.

Set	In-distribution sources	Tr/Val/Cal/Test	Distribution-shift probe
D1	Breast (3; 781) ( Al-Dhabyani et al., 2020 ) ; kidney (2; 9416) ( imtkaggleteam, n.d. ) ; liver fibrosis (5; 6323) ( Joo et al., 2023 ) . Total: 10 classes, 16520 images.	10737 / 1651 / 828 / 3304	Lung ultrasound, 2149 images ( Born et al., 2020 ; Madhu et al., 2024 ) .
D2	Breast (2; 927) ( Huang et al., 2023 ) ; kidney (2; 540) and pelvic (2; 625) ( Karim, 2025 ) ; ovary (2; 1924) ( amritarajput54, n.d. ) ; thyroid (2; 387) ( Pang et al., 2025 ) . Total: 10 classes, 4403 images.	2864 / 440 / 219 / 880	Fetal ultrasound, 1669 images ( Anitha, 2024 ) .

Table 2: Composition and evaluation roles of the two image-level ultrasound collections. Counts in the source column are shown as number of classes and images. Tr/Val/Cal/Test denotes the frozen 65/10/5/20% partition.

Figure 2: Datasets and frozen audit protocol. Per-class image counts for (a) Dataset 2 (4403 images, five organ groups) and (b) Dataset 1 (16520 images, three organ groups), coloured by organ. The shading within each bar shows the frozen 65/10/5/20 image-level partition of that class. (c) The shared protocol: fixed partition seed, ten matched training seeds, paired two-sided 90%

t

intervals (one-sided 95% decisions), the prespecified margins, and the fetal and lung OOD probes, which contribute only to the separate OOD decision.

Classifier	Seeds	Params (M)	MACs (M)	Macro-F1	TS-ECE	TS-NLL
MobileNetV3-Small–FC	3	0.933	54.9	$0.785\pm 0.004$	0.036	0.392
ResNet18–FC	3	11.182	1813.6	$0.802\pm 0.011$	0.029	0.377

Table 3: Conventional pretrained-CNN context on Dataset 2. Values are mean

\pm

sample standard deviation over three matched seeds.

Figure 3: Computational cost versus task performance on Dataset 2. Raw macro-F1 (mean

\pm

SD across seeds) is plotted against multiply–accumulate operations on a logarithmic axis. Marker area is proportional to parameter count, and the dashed line connects the nondominated mean operating points. The shaded band marks the compact operating range. The two contextual baselines used three seeds and did not enter formal decisions. Descriptively, ResNet18 required approximately fourteen times the computation of the audited candidates for a

+0.012

mean macro-F1 difference over Full-EDL.

Figure 4: Raw-MSP risk–coverage curves for (a) Dataset 2 and (b) Dataset 1. Curves show the mean selective risk across the ten matched seeds as coverage varies under confidence-ranked referral. Shaded bands show

\pm

1 SD for Simple-CE and Full-EDL, and thin grey lines show the remaining candidates. Legend values report mean raw-MSP AURC. Simple-CE maintained lower mean selective risk than Full-EDL over most of the coverage range on both datasets. These curves are supporting descriptive evidence, not formal decision criteria.

Candidate	Macro-F1 (F1)	Raw ECE	TS ECE	Raw NLL	TS-NLL (NLL)	Raw-MSP AURC	Raw $1-\mathrm{MSP}$ OOD AUROC (OOD)
Full-EDL	$0.790\!\pm\!0.017$ /Ref.	$0.051\!\pm\!0.015$	$0.039\!\pm\!0.008$	$0.437\!\pm\!0.022$	$0.427\!\pm\!0.024$ /Ref.	$0.083\!\pm\!0.011$	$0.960\!\pm\!0.027$ /Ref.
Full-CE	$0.789\!\pm\!0.011$ /I	$0.085\!\pm\!0.025$	$0.029\!\pm\!0.007$	$0.480\!\pm\!0.086$	$0.373\!\pm\!0.027$ /P	$0.052\!\pm\!0.011$	$0.948\!\pm\!0.016$ /NV
Simple-EDL	$0.778\!\pm\!0.024$ /I	$0.057\!\pm\!0.015$	$0.044\!\pm\!0.011$	$0.457\!\pm\!0.034$	$0.443\!\pm\!0.034$ /I	$0.088\!\pm\!0.014$	$0.961\!\pm\!0.013$ /NV
Simple-CE	$0.783\!\pm\!0.015$ /I	$0.078\!\pm\!0.016$	$0.031\!\pm\!0.006$	$0.455\!\pm\!0.066$	$0.376\!\pm\!0.019$ /P	$0.051\!\pm\!0.007$	$0.936\!\pm\!0.019$ /V
Expert-only-EDL	$0.785\!\pm\!0.013$ /I	$0.060\!\pm\!0.023$	$0.041\!\pm\!0.010$	$0.447\!\pm\!0.025$	$0.432\!\pm\!0.020$ /P	$0.088\!\pm\!0.011$	$0.961\!\pm\!0.015$ /NV
Slim-CE	$0.789\!\pm\!0.022$ /I	$0.077\!\pm\!0.013$	$0.025\!\pm\!0.005$	$0.447\!\pm\!0.070$	$0.368\!\pm\!0.023$ /P	$0.049\!\pm\!0.010$	$0.940\!\pm\!0.015$ /V

Table 4: Dataset 2 primary controlled-audit results. Values are mean

\pm

sample standard deviation across ten matched seeds. ECE and NLL are shown as raw/temperature-scaled (TS) values.

Figure 5: Paired candidate-minus-Full-EDL effects on (left) macro-F1, (centre) temperature-scaled negative log-likelihood (TS-NLL), and (right) primary raw

1-\mathrm{MSP}

OOD AUROC for (a) Dataset 2 and (b) Dataset 1. Points show paired mean differences and horizontal bars show two-sided 90% paired

t

intervals across ten matched seeds. Grey dashed lines mark zero. Red dashed lines mark the prespecified F1 (

-0.01

), TS-NLL (

+0.02

), and OOD-veto (

-0.03

) margins. Filled markers denote PASS , open markers denote INCONCLUSIVE or NOT_VETOED , and red-edged downward triangles denote VETO . Interval colour identifies each candidate consistently across figures. Decisions are interpreted separately for each endpoint and dataset.

Figure 6: Symmetric post-hoc calibration audit for (a) Dataset 2 and (b) Dataset 1. Each candidate is plotted by ECE and NLL. Lower values on both axes are preferable. Open markers show raw outputs, filled markers show the corresponding temperature-scaled outputs, and arrows connect the two states for each candidate. Faint whiskers show two-sided 90%

t

intervals for the temperature-scaled state across ten seeds. Numbers identify candidates as in Fig. 3 . The apparent raw-calibration advantage of EDL on Dataset 2 was not retained after symmetric temperature scaling and did not recur on Dataset 1.

Candidate	Macro-F1 (F1)	Raw ECE	TS ECE	Raw NLL	TS-NLL (NLL)	Raw-MSP AURC	Raw $1-\mathrm{MSP}$ OOD AUROC (OOD)
Full-EDL	$0.912\!\pm\!0.039$ /Ref.	$0.057\!\pm\!0.023$	$0.013\!\pm\!0.005$	$0.170\!\pm\!0.098$	$0.137\!\pm\!0.090$ /Ref.	$0.013\!\pm\!0.012$	$0.920\!\pm\!0.068$ /Ref.
Simple-EDL	$0.927\!\pm\!0.014$ /I	$0.048\!\pm\!0.008$	$0.013\!\pm\!0.004$	$0.137\!\pm\!0.021$	$0.109\!\pm\!0.023$ /I	$0.010\!\pm\!0.002$	$0.930\!\pm\!0.020$ /NV
Simple-CE	$0.923\!\pm\!0.013$ /I	$0.017\!\pm\!0.004$	$0.009\!\pm\!0.003$	$0.121\!\pm\!0.030$	$0.087\!\pm\!0.018$ /P	$0.003\!\pm\!0.001$	$0.933\!\pm\!0.019$ /NV

Table 5: Dataset 1 second-dataset internal replication. Values are mean

\pm

sample standard deviation across ten matched seeds. ECE and NLL are shown as raw/temperature-scaled (TS) values.

Figure 7: Per-class F1 (mean over the ten matched seeds) for (a) the six Dataset 2 candidates and (b) the three Dataset 1 replication candidates. Within each dataset, the per-class profiles were similar across candidates. The dagger marks a class with reference support below 30 images, reported descriptively under the prespecified low-support rule.

Figure 8: Cross-dataset synthesis and restricted transfer boundary. (a) Evidence matrix: the upper band collects findings repeated on both datasets (no established macro-F1 gain from added complexity, and CE+TS calibration matched or better), and the lower band collects dataset-specific findings (the EDL raw-ECE advantage and the Simple-CE OOD veto, marked by the red triangle). Open markers denote patterns not observed. (b) Restricted zero-shot breast stress test: ten Dataset 2-trained Full-EDL checkpoints applied without target fine-tuning or temperature refitting to 648 Dataset 1 breast images. Coarse organ assignment was largely retained (0.48% cross-organ error), whereas benign/malignant discrimination (

0.59\pm 0.04

macro-F1 relative to the plotted 0.5 reference) and source-domain calibration transfer (ECE 0.18 to 0.19) remained limited.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: Complete Full-EDL maximal-candidate architecture. MobileNetV3-Small is divided into low- and high-level stages. NALA performs intermediate spatial-reliability weighting. The auxiliary evidential probe, evidential router, four expert blocks, shared fallback, and parameter-free reliability gate form the adaptive branch. The gated branch is combined with the high-level backbone feature through a positive residual scale before global average pooling and evidential output. Full-EDL is shown as the maximal reference evaluated in the architecture audit.

Figure A.2: Computational structure of the NALA spatial-reliability module. The attention branch combines ReLU- and ELU-based interactions between projected query and key features. A separate label-free branch produces a learned suppression map. Their product forms a broadcast spatial-reliability map that modulates only the positive-scaled residual contribution, leaving the identity path intact. The suppression map is an internal weighting signal and is not treated as a verified physical noise or artefact estimate.

Figure A.3: Detailed UGMoE routing, expert mixing, and reliability-gating computation. An evidential router produces sparse top-2 expert weights and router uncertainty. Four expert outputs are mixed and interpolated with a shared fallback by a parameter-free gate formed from routing-confidence, probe-uncertainty, and router-uncertainty factors. The resulting adaptive output is added to the backbone feature through a positive residual scale. Dashed lines denote uncertainty or control signals. Solid lines denote data flow.

Explore similar work

Oct 9, 2026

C $^2$ LUE: Clinically Aligned Patient-Level Ultrasound Diagnosis via Evidence-Aware Multi-Instance Learning

Patient-level ultrasound diagnosis demands not only high predictive accuracy but also clinically aligned mitigation of asymmetric diagnostic risks. While Multi-Instance Learning (MIL) effectively models routine multi-view exams, standard aggregators often dilute heterogeneous diagnostic cues and optimize solely for global accuracy. This frequently yields undesirable sensitivity-specificity trade-offs. To bridge this gap, we propose C

^2

LUE, a novel MIL framework coupling evidence aware aggregation with risk-aligned regularization. To preserve evidence heterogeneity and prevent unimodal attention collapse, C

^2

LUE employs a multi-query cross-attention mechanism built upon a LoRA-adapted DINOv3 backbone. Multiple latent queries explicitly extract and disentangle diverse morphological patterns from the ultrasound sequence. Furthermore, to improve clinical risk alignment, we introduce a dynamic trade-off regularization objective that penalizes sensitivity-specificity imbalance during training. Rather than leaving the clinical trade-off entirely to post-hoc threshold adjustment, this objective encourages the learned score distribution toward a more balanced operating region. Extensive evaluations on patient-level, biopsy-proven thyroid and breast ultrasound datasets demonstrate that C

^2

LUE consistently outperforms prior alternatives. By uniting strong discrimination with a balanced error profile, C

^2

LUE provides a promising solution for real-world ultrasound screening and diagnosis.

Qin-Yi Zhang, Linying Cao, De-Xing Huang +7

State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China

Multiple Instance Learning Ultrasound Imaging Classification

Ask PDF

Sep 17, 2026

Uncertainty-driven training for three-dimensional calibrated lung nodule classification

In this work, we present an uncertainty-driven training framework for three-dimensional computed tomography (CT) lung nodule classification, where validation-based uncertainty estimates guide loss reweighting to enhance predictive performance and probability calibration. Two Uncertainty Quantification (UQ) methods are considered: Monte Carlo Dropout (MCD) and Evidential Deep Learning (EDL). Both provide per-class uncertainty estimates that modulate the loss and encourage focus on hard or unreliable classes. The framework is evaluated with ResNet, DenseNet, EfficientNet, Vision Transformer (ViT), and Swin Transformer backbones on two datasets: the clinical LIDC-IDRI cohort and the NoduleMNIST3D benchmark. Uncertainty-driven training achieves classification performance similar to conventional training while substantially improving calibration, with an expected calibration error (ECE) reduced by up to 65% on LIDC-IDRI. EDL attains competitive performance on shallower architectures with single-pass inference, whereas MCD is more robust on deeper networks. Analysis across architectural families reveals that uncertainty-driven training benefits convolutional backbones more consistently than transformer-based architectures: EDL in particular degrades on ViT, suggesting that the Dirichlet evidence parameterisation may interact unfavourably with attention-based architectures at lower input resolutions. A posteriori temperature scaling proves highly effective across all configurations, indicating that a simple scalar calibration can be competitive even without explicit uncertainty-aware training. Our results indicate that integrating UQ into the training loop can significantly improve probabilistic calibration and support more trustworthy deployment of three-dimensional medical imaging models.

Giuseppe Tripodi, Alessandro De Rosis, Saleh Rezaeiravesh

School of Engineering, The University of Manchester, Manchester (UK)

Medical Image Classification Uncertainty Calibration Uncertainty Quantification 3D Medical Imaging Evidential Deep Learning

Ask PDF

Apr 30, 2026

Echo-α: Large Agentic Multimodal Reasoning Model for Ultrasound Interpretation

Ultrasound interpretation requires both precise lesion localization and holistic clinical reasoning, yet existing methods typically excel at only one of these capabilities: specialized detectors offer strong localization but limited reasoning, whereas multimodal large language models (MLLMs) provide flexible reasoning but weak grounding in specialized medical domains. We present Echo-α, an agentic multimodal reasoning model for ultrasound interpretation that unifies these strengths within an invoke-and-reason framework. Echo-α is trained to coordinate organ-specific detector outputs, integrate them with global visual context, and convert the resulting evidence into grounded diagnostic decisions beyond detector-only inference. This behavior is established through a nine-task supervised curriculum and then refined by sequential reinforcement learning under different reward trade-offs, yielding Echo-α-Grounding for lesion anchoring and Echo-α-Diagnosis for final diagnosis. On multi-center renal and breast ultrasound benchmarks, Echo-α outperforms competitive baselines on both grounding and diagnosis. In particular, on cross-center test sets, Echo-α-Grounding attains 56.73%/43.78% [email protected] and Echo- α-Diagnosis reaches 74.90%/49.20% overall accuracy on renal/breast ultrasound. These results suggest that agentic multimodal reasoning can turn specialized detectors into verifiable clinical evidence, offering a practical route toward ultrasound AI systems that are more accurate, interpretable, and transferable. The repository is at https://github.com/MiliLab/Echo-Alpha.

Jing Zhang, Wentao Jiang, Tao Huang +8

School of Computer Science, Wuhan University, China · The Central Hospital of Wuhan, China · Taizhou Hospital of Zhejiang Province, China +1

Medical Image Analysis Multimodal Reasoning Ultrasound Imaging Medical VLMs Multimodal Agents Medical Image Localization

Ask PDF

Abstract

Explore similar work

C2^22LUE: Clinically Aligned Patient-Level Ultrasound Diagnosis via Evidence-Aware Multi-Instance Learning

Uncertainty-driven training for three-dimensional calibrated lung nodule classification

Echo-α: Large Agentic Multimodal Reasoning Model for Ultrasound Interpretation

C $^2$ LUE: Clinically Aligned Patient-Level Ultrasound Diagnosis via Evidence-Aware Multi-Instance Learning