Conditional Medical Image Routing Framework Balances Diagnostic Accuracy and Computational Latency

A retrospective investigation published in Discover Computing has demonstrated that a hierarchical conditional inference architecture can effectively balance medical image classification performance with computational efficiency by triaging diagnostic cases according to predictive confidence. Conventional deep learning implementations in healthcare informatics typically deploy monolithic models that channel all diagnostic imagery through identical computational pathways, regardless of whether individual scans represent routine clinical presentations or ambiguous pathologies. To resolve this operational inefficiency, an engineering research team designed Computer-Aided Triage and Diagnostic Routing, abbreviated as CATDR, which integrates a lightweight primary screening model with an automated routing mechanism that selectively escalates uncertain cases to specialized downstream neural networks. When evaluated across three benchmark datasets spanning fetal ultrasound, gastrointestinal endoscopy, and dermoscopic skin lesions, the proposed framework achieved classification accuracies of 94.40 percent, 93.10 percent, and 73.80 percent, while maintaining average inference latencies of 4.60 milliseconds, 4.92 milliseconds, and 14.51 milliseconds per image. The authors noted that these initial findings provide technical feasibility evidence for confidence-guided diagnostic escalation rather than immediate prospective clinical validation.

Clinical decision support tools face strict operational constraints because diagnostic demands continue to rise while specialized healthcare resources remain strictly bounded across hospital networks. While standard benchmark evaluations often prioritize top-tier classification accuracy above all other parameters, practical clinical adoption requires deep learning models to satisfy workflow compatibility, operational reliability, and efficient resource allocation. In routine diagnostic screening environments, visually distinct and canonical presentations constitute a substantial proportion of incoming examinations, making exhaustive computational processing across entire network backbones redundant for straightforward samples. However, early-exit networks remain fundamentally constrained within a single parameter space, preventing the dynamic redirection of ambiguous cases toward distinct, specialized neural architectures uniquely structured for difficult diagnostic interpretations.

The CATDR framework addresses these structural limitations through a decoupled three-tier architecture that explicitly separates initial case screening, algorithmic expert assignment, and multi-depth prediction selection. The primary tier consists of an EfficientNet-B0 convolutional network, selected for its balanced compound scaling of depth, width, and resolution, which processes input images resized to 224 by 224 pixels. Applying a normalized softmax transformation converts these raw outputs into an operational score distribution, from which the framework extracts the maximum class probability as a functional screening confidence metric. Under the primary experimental protocol, a predefined decision threshold of 0.70 dictates whether inference terminates immediately or advances into downstream evaluation. Cases meeting or exceeding this confidence boundary are finalized instantly at the screening tier, thereby bypassing all subsequent computation, whereas cases falling below the threshold are packaged with their intermediate feature embeddings and redirected to the second operational tier.

The secondary routing tier functions as an intelligent computational dispatcher that activates strictly for ambiguous cases that fail the screening confidence threshold. Rather than reloading or recomputing the raw diagnostic image from scratch, the routing engine ingests a concatenated vector comprising the intermediate screening feature representation and the full softmax probability distribution. This unified vector is processed by a dedicated multi-layer perceptron equipped with a 256-dimensional hidden layer that computes binary assignment logits to designate the optimal downstream specialist. To circumvent the prohibitive burden of acquiring manual, expert-specific annotations for router training, the researchers engineered a complementarity-driven pseudo-labeling strategy applied across the development data split. Under this training paradigm, retained low-confidence development samples were evaluated by two downstream expert models, and pseudo-supervision targets were assigned exclusively when one expert model correctly classified an image while the alternative expert failed. Development instances where both experts agreed in correctness or error were systematically excluded to avoid ambiguous assignment signals, allowing the router to learn specialized algorithmic escalation policies directly from model complementarity.

The third tier of the system incorporates two structurally distinct downstream expert architectures, specifically ResNet50 designated as Expert A and EfficientNet-B3 designated as Expert B. Each expert model is augmented with three auxiliary classification heads positioned at progressive architectural depths, designed to deliver depth-specific categorical predictions for escalated cases. During inference, intermediate confidence scores from the shallower output heads are sequentially evaluated against strict preset decision boundaries, which were fixed at 0.96 for Head 1 and 0.98 for Head 2 across both models. If neither shallow head satisfies its respective confidence threshold, the framework automatically adopts the classification generated by the deepest terminal head, Head 3. The investigators specifically underscored that all three auxiliary heads are computed sequentially within a forward pass, meaning that this mechanism functions as a logical prediction-selection rule rather than a physical runtime early-exit mechanism. Consequently, output selection does not truncate latency within the selected expert network itself, but ensures that predictions are drawn from internal representational levels exhibiting adequate diagnostic certainty.

Rigorous experimental evaluation was conducted on three public medical imaging repositories representing diverse clinical imaging modalities and task complexities. Input images were uniformly resized, and models were optimized using AdamW with an initial learning rate of 0.0001, a weight decay of 0.0001, and a batch size of 16 across up to 20 training epochs with early stopping patience set to five epochs. For the Kvasir v2 and PAD-UFES-20 datasets, stratified partitions established a 64 percent training, 16 percent validation, and 20 percent held-out testing division, while Fetal-Planes DB followed official splits with a 20 percent validation holdout derived from the training partition. Hardware benchmarking was executed in PyTorch on an NVIDIA RTX 4090 GPU paired with an Intel Xeon Platinum 8352V processor, recording latencies across multiple runs with full synchronization.

Quantitative benchmarking against five static, single-path deep learning baselines demonstrated the diagnostic robustness of the conditional triage strategy across heterogeneous imaging domains. In primary held-out testing, CATDR attained higher numerical accuracy and Cohen’s kappa values than standalone EfficientNet-B1, ResNet50, ResNet18, ShuffleNet V2, and MobileNetV3-S across the fetal ultrasound and endoscopy datasets. A complementary paired-bootstrap uncertainty analysis on matched sample predictions revealed that diagnostic gains over the standalone EfficientNet-B0 screening backbone were statistically confirmed on Fetal-Planes DB, where confidence intervals for accuracy, F1-score, sensitivity, and kappa strictly excluded zero. Conversely, on Kvasir v2 and PAD-UFES-20, bootstrap intervals crossed zero, indicating that while numerical improvements were observed, they represent modest empirical gains rather than established statistical superiorities.

The core computational efficiency of the framework stems directly from its capacity to prevent unnecessary downstream model invocations for the vast majority of routine diagnostic cases. Analysis of triage behavior revealed that CATDR successfully finalized 98.74 percent of fetal ultrasound scans and 96.71 percent of endoscopic images at the initial EfficientNet-B0 screening stage, escalating only 1.26 percent and 3.29 percent of cases to downstream experts. Even on the challenging PAD-UFES-20 skin lesion collection, 86.81 percent of samples were resolved during primary screening, with only 13.19 percent requiring expert routing. Compared to a static ResNet50 baseline, CATDR reduced average inference latency by 64.01 percent on Fetal-Planes DB, 68.16 percent on Kvasir v2, and 63.10 percent on PAD-UFES-20. When contrasted against EfficientNet-B1, latency reductions reached 38.42 percent, 36.02 percent, and 31.43 percent across the respective datasets. In a dedicated unified timing protocol evaluating single-image processing, CATDR recorded latencies of 0.245 milliseconds on ultrasound and 0.248 milliseconds on endoscopy, illustrating that conditional triage introduces negligible overhead over standalone screening while shielding clinical infrastructure from excessive computational demand.

Because the gating mechanism relies fundamentally on maximum softmax probabilities, the researchers conducted detailed calibration assessments to ensure that screening confidence reliably mirrored predictive accuracy. Evaluating reliability diagrams and calibration metrics demonstrated that applying temperature scaling fitted exclusively on development data substantially improved probability calibration across all datasets. Expected calibration error declined from 0.0420 to 0.0128 on Fetal-Planes DB, from 0.0376 to 0.0192 on Kvasir v2, and from 0.1578 to 0.0368 on PAD-UFES-20. Safety analysis at the standard 0.70 threshold revealed that 255 ultrasound, 84 endoscopy, and 82 skin lesion samples were incorrectly finalized at screening, corresponding to finalized error rates of 4.96 percent, 5.48 percent, and 21.98 percent. Adjusting the screening threshold upward to 0.85 on PAD-UFES-20 reduced incorrectly finalized instances to 21, but dropped screening coverage to 38.04 percent, underscoring that threshold configuration involves an unavoidable operational trade-off between coverage volume and diagnostic risk.

The study authors documented several critical methodological limitations that must govern the interpretation of their findings prior to any prospective implementation. Grouping metadata audits established that the benchmark partitions for Fetal-Planes DB and PAD-UFES-20 do not guarantee strict patient-level or lesion-level independence, while Kvasir v2 entirely lacks patient identifiers, meaning that reported metrics reflect benchmark-specific behavior rather than verified generalizability across independent patient populations. Furthermore, because the routing module depends on training samples where downstream experts disagree in diagnostic accuracy, pseudo-supervision signals proved highly sparse, comprising only tiny evaluation subsets of two, four, and six samples across the respective datasets. Retrospective oracle-routing analyses confirmed that substantial unrealized complementarity remained untapped, with theoretical oracle accuracy gains reaching 4.565 percentage points on PAD-UFES-20. Additionally, maintaining a multi-tiered architecture increases training complexity and system maintenance requirements compared to monolithic baselines. The authors concluded that while CATDR demonstrates the technical feasibility of dynamic medical triage, prospective clinical utility will depend on rigorous patient-independent validations, adaptive threshold calibration, and expanded routing supervision strategies.

Subject of Research: Conditional inference and diagnostic routing algorithms for medical image classification

Article Title: Computer-aided triage and diagnostic routing for efficient medical image classification

Article References: Pan, W., Wang, S., Ma, Z., Lv, C., & Wang, Y. (2026). Computer-aided triage and diagnostic routing for efficient medical image classification. Discover Computing, 29(1), Article 799. https://doi.org/10.1007/s10791-026-10695-9

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10695-9

Keywords: medical image classification, conditional inference, computer-aided triage, diagnostic routing, EfficientNet-B0, ResNet50, EfficientNet-B3, temperature scaling, expected calibration error, selective prediction, fetal ultrasound, gastrointestinal endoscopy, dermoscopy, Fetal-Planes DB, Kvasir v2, PAD-UFES-20, learning to defer, multi-tier deep learning