Bigger EfficientNet Models Did Not Consistently Improve Retinal Scan Classification

A controlled comparison, with an unresolved data caveat

A November 2026 preprint by Shkëlqim Sherifi and colleagues compares all eight EfficientNet variants, B0 through B7, for classifying spectral-domain optical coherence tomography images. These scans provide cross-sectional views of retinal structure. Under the study’s training and evaluation protocol, larger networks did not consistently achieve higher accuracy.

The strongest qualification concerns the data. The researchers retained the downloaded training, validation, and test partitions without independently verifying patient-level separation or checking for duplicate images across partitions. Consequently, they cannot establish that test-image content was absent from training. The reported scores are conditional on that supplied split, rather than confirmed performance on independent image content.

What the experiment held constant

EfficientNet’s original scaling approach increases network depth, width, and image resolution together. This study deliberately kept every input at 224 × 224 pixels while changing the model variant. It therefore examines architectural scaling at fixed resolution, not the complete original scaling strategy.

The researchers used a Kaggle-distributed release of the Kermany retinal OCT dataset containing 109,309 images. Its four categories were choroidal neovascularization, diabetic macular edema, Drusen, and Normal. The supplied partitions contained 76,515 training images, 21,861 validation images, and 10,933 test images. Normal was the largest category and Drusen the smallest.

All models began with ImageNet-pretrained weights. Each grayscale scan was copied into three identical channels to match the backbones’ expected input. The models shared an augmentation policy, classification-head design, and two-stage training procedure. First, the pretrained backbone stayed frozen while the new classifier trained. Then approximately its upper half was unfrozen for joint fine-tuning. Class weighting addressed the imbalance during optimization, and validation loss determined checkpoint selection.

A smaller variant led the accuracy comparison

Reported test accuracy ranged from 95.93% for B0 to 96.95% for B1. Accuracy fluctuated across the intermediate variants before reaching 96.73% for B7, the largest model. Thus, B1 exceeded B7 by just 0.22 percentage points in this experiment.

That ordering differed from the ImageNet results cited in the paper, where B7 ranked highest. The comparison suggests that the source benchmark’s ordering did not transfer unchanged under this particular OCT protocol. It does not establish that larger EfficientNet models are generally worse for retinal imaging.

The qualification matters because every variant was evaluated with only one nominal random seed. The authors explicitly note that this does not quantify run-to-run variability or establish whether closely matched models would retain their ordering across independent repetitions.

Drusen exposed differences hidden by overall accuracy

Across every model, the class-level F1 ordering was Normal, then choroidal neovascularization, diabetic macular edema, and finally Drusen. F1 summarizes the balance between precision and recall: how reliably Drusen predictions identified that category, and how many actual Drusen images were detected.

B1 achieved the highest reported Drusen F1, 0.8780, with precision of 0.8493 and recall of 0.9087. B7 detected a greater proportion of Drusen images, achieving the highest recall, 0.9177. However, its lower precision of 0.8306 left its F1 slightly below B1’s, at 0.8720.

Drusen accounted for 82.2% of the summed squared deviations in class-specific F1 across the eight models. This describes where variation between models was concentrated; it is not the percentage of all classification errors attributable to Drusen. The result illustrates why aggregate accuracy alone gives an incomplete account of this comparison.

Computational cost rose more clearly than accuracy

B1 contained 7.37 million total parameters, compared with 65.54 million for B7. Their measured mean CPU inference times were approximately 513 and 1,248 milliseconds per image, respectively. These measurements used batch size one on an Intel Core i7-6700HQ, after warm-up runs.

The machine was not dedicated to benchmarking, so background activity could affect timing. The values describe that hardware and software environment, not universal deployment speeds. Likewise, the paper’s accuracy-to-parameter ratio is a descriptive comparison, not a general efficiency law accounting for energy use or other hardware constraints.

What remains unresolved

The authors identify a published duplicate-image audit of an OCT dataset variant with matching partition counts. However, they did not compare their own files against the audited files, so those published overlap counts cannot be assigned directly to this experiment. A file-level audit and, where necessary, leakage-controlled retraining remain essential.

The study also lacks external-dataset validation and does not isolate the effects of augmentation, resolution, or fine-tuning choices. Its attribution maps and embedding visualizations are qualitative, not causal explanations of B1’s performance. The preprint supports a bounded conclusion: evaluate model size directly on the target task, while treating this particular ranking as provisional.