Abstract

Medical imaging models are usually benchmarked on large, curated datasets and evaluated on hardware that most regional clinics do not have. This review examines convolutional architectures for medical image analysis under a different set of constraints: limited annotated data, no dedicated GPU, and no on-site radiologist to validate borderline cases.

---

Motivation

Diagnostic imaging is one of the areas where deep learning has shown the most consistent results. Yet the deployment gap remains wide: the clinics that would benefit most from decision support are precisely those least able to run it.

We identify four recurring constraints:

• Annotated datasets in the order of hundreds, not hundreds of thousands, of images.
• Inference hardware limited to CPU or entry-level accelerators.
• Intermittent connectivity, which rules out permanent cloud inference.
• Absence of a specialist able to audit model output on site.

---

Scope of the review

We surveyed work published between 2019 and 2024 addressing three imaging modalities commonly available in regional facilities:

Chest radiography

The most widely available modality. Literature concentrates on pneumonia, tuberculosis and cardiomegaly screening.

Dermatological photography

Attractive because acquisition requires only a smartphone camera, but highly sensitive to lighting and skin-tone distribution in the training set.

Ocular fundus imaging

Mainly applied to diabetic retinopathy grading, with the strongest evidence of clinical utility among the three.

---

Findings

Transfer learning dominates

Nearly all reviewed work under small-data conditions relies on backbones pretrained on natural images. Fine-tuning only the final blocks consistently outperformed training from scratch when fewer than 2 000 annotated samples were available.

Model compression is underreported

Although quantisation and distillation reduce inference cost substantially, fewer than a third of the reviewed papers report latency or memory footprint at all — the metrics that determine whether a model is deployable in the target setting.

Distribution shift is the main failure mode

Models trained on one institution’s equipment degraded noticeably when applied to images from different devices. Reported drops in balanced accuracy ranged from 8 to 21 percentage points across the studies that evaluated external validation.

Calibration matters more than raw accuracy

In the absence of an on-site specialist, an overconfident false negative is more damaging than a lower-accuracy model that reliably flags its own uncertainty. Very few reviewed systems report calibration metrics.

---

Discussion

The literature optimises for the wrong target when the deployment context is a low-resource clinic. Headline accuracy on a single-institution test split says little about behaviour on a different X-ray machine, and nothing about whether the model will run on the available hardware.

We argue that work aimed at these settings should report, as a minimum: external validation on at least one independent source, inference cost on commodity hardware, and a calibration measure alongside accuracy.

---

Conclusions

• Transfer learning remains the most reliable strategy under small-data constraints.
• Distribution shift across acquisition devices is the dominant obstacle to deployment.
• Calibration and referral behaviour deserve more attention than incremental accuracy gains.
• Reporting practices should include deployment-relevant metrics, not only predictive performance.

---

Future work

• Establish a shared benchmark reflecting realistic acquisition variability.
• Evaluate distillation pipelines targeting CPU-only inference.
• Study referral workflows where the model defers uncertain cases rather than predicting.