Md. Asif Uddin

Plate I

HierarchiRetina

HierarchiRetina: A Lesion-Aware Hierarchical Deep Learning Framework Enabling Interpretable Diabetic Retinopathy Severity Assessment via Multi-Stage Retinal Fundus Image Analysis

Deposited
Year
2026
Supervisors
Rafeed Rahman, Niloy Farhan, Md. Golam Rabiul Alam
Co-authors
Hosni Rabbani, Shahrin Tabassum
Datasets
EyePACS · APTOS 2019 · Messidor-2 · IDRiD · DDR · FGADR · DRIVE · STARE · MAPLES-DR · HRF · E-Ophtha
Abstract

Diabetic Retinopathy (DR) is a leading cause of preventable blindness. It affects millions of people with diabetes, especially where eye care is hard to reach. Most deep learning systems treat DR assessment as a single end-to-end classification task. This approach does not reflect how the disease is handled in the clinic. It blends screening with severity staging and it produces opaque predictions that doctors find hard to trust. We present HierarchiRetina, a three-stage framework that mirrors the way a clini- cian works. Stage I uses a ConvNeXt V2 Large model to quickly separate healthy retinas from diseased ones. Stage II runs only on diseased images. It produces pixel- level masks for four lesion types, microaneurysms (MA), haemorrhages (HE), hard exudates (EX) and cotton-wool spots (CWS), along with the retinal vessels. Stage III is LG-DRG (Lesion-Guided DR-Grader). It fuses the original image with the five masks and uses two heads: one for severity (Mild, Moderate, Severe, Proliferative) and one for gradability, which routes unreadable images to a human reviewer. Three design choices set this work apart. The lesion model uses experts of different types, not copies of one block, so each lesion shape is handled by a specialist filter. The grader lets the masks guide attention through a gate that starts switched off, so the model learns to trust the masks only once they become reliable. The dual-head split moves unreadable images off the severity scale instead of mis-ranking them on it. We evaluate HierarchiRetina on six public datasets at once (EyePACS, DDR, AP- TOS 2019, Messidor-2, IDRiD, FGADR), under a strict end-to-end rule that counts every screening miss as a final grading error. Even with a single early-stopped fold and this harder protocol, our system sets a new state-of-the-art on the official DDR test split, surpassing the strong published FA+KC-Net+R2 benchmark reported by Tian and colleagues [50]. The lesion masks also make the system interpretable: a clinician can see which lesions were detected, where they are and how they shaped the final grade. Together, these results show that a staged, lesion-aware design can match clinical reasoning, deliver strong grading accuracy and earn the trust needed for real-world screening.

[SUPPLY]

Results

MeasureValueNote
Macro-AUC0.9422
Ungradable detection AUC0.9995
DDR 5-class QWK, 5-fold ensemble0.8533n = 3,759; published benchmark on the same split, 0.8617
End-to-end pipeline QWK0.7942
Per-dataset QWK0.83 – 0.93FGADR to APTOS
The HierarchiRetina three-stage pipelineA fundus image enters Stage I screening for referable diabetic retinopathy at grade 2 or above. It then branches to five parallel lesion segmentation models — microaneurysms, haemorrhages, hard exudates, cotton wool spots and vessels — whose masks rejoin the raw image as eight channels into Stage III, the lesion-guided grader. Stage III emits either an ordinal severity grade from one to four, or an ungradable verdict routed to human review.Fundus imageStage I · screening(referable DR, grade ≥ 2)MAHSMoEHEMultiScaleEXBrightSpotCWSwarm-upVesselsFOV maskStage III · LG-DRG(8 channels: RGB + 5 masks)Severity grade(CORN ordinal, 1–4)Ungradable(routed to human review)
Fig. I — The three-stage pipeline. Screening gates the image; five lesion models segment in parallel, each in its own colour; their masks rejoin the raw image as eight channels for grading. Ungradable images leave the pipeline rather than receive a grade.

The decision that set the number

DDR grades diabetic retinopathy on a five-point scale, and grade 5 is not a severity at all. It means the photograph cannot be read — the media are opaque, the field is wrong, the exposure failed. Treating it as an ordinal level above proliferative disease is arithmetically convenient and clinically false.

Left as an ordinal class, grade 5 pushes QWK to 0.9017. That number is not comparable to anything, because published work on the same split does not treat gradability as a severity. Splitting it out into a separate binary gradability head produced 0.8533 — lower, honest, and comparable to the 0.8617 benchmark on the identical split.

It is also the better clinical design. An ungradable photograph should be routed to a human, not assigned a fabricated grade. The pipeline now says *I cannot read this* instead of guessing, and that is a different and more useful failure.

Stage I · Screening

Referable disease is grade ≥ 2. Two backbones carry this stage: SwinV2-Large at 384px and ConvNeXtV2-Large at 512px, trained on roughly 56,455 DDR images.

Validation AUC reached 0.9475; held-out test AUC was 0.9043 across 58,689 images. The operating threshold was calibrated on validation only and locked before the test set was read once.

Stage II · Five lesions, five designs

A microaneurysm is a handful of dark pixels. A cotton wool spot is a soft pale patch with no edge worth speaking of. They do not respond to the same architecture, and they do not respond to the same loss.

Each lesion class gets its own model, its own loss composition, and its own sampling regime — microaneurysms and haemorrhages through the sparse-lesion designs, hard exudates through bright-spot specialisation, vessels through the high-resolution hybrid under a field-of-view mask.

The cotton wool spot collapse

Version 1 failed completely: val_loss NaN, Dice 0.0, precision 1.0, recall 0.0. A model predicting nothing scores perfect precision, and that is the tell.

Three causes, all real:

  • TextureExpert computed local variance as E[x²] − E[x]². Under mixed-precision training that subtraction overflows in FP16 and propagates NaN through the whole graph.
  • RegionCoherenceLoss used plain total variation. The global minimum of total variation is the all-zero prediction. The model was not broken — it was correctly minimising a loss that paid it to predict nothing.
  • No loss warmup and no pos_weight on an extremely sparse class.

Version 2 fixed all three: mean absolute deviation in place of variance for FP16 safety; probability-weighted total variation that returns zero loss on zero predictions; a three-phase loss warmup; NaN guards in both the training and validation loops; gradient clipping at 0.5; learning rate down to 7e-5.

Stage III · Lesion-guided grading

The grader takes eight channels — RGB plus the five lesion masks — into a ConvNeXtV2-Large with cross-attention lesion gating. The severity head is a CORN ordinal regressor over grades 1–4. Gradability is a separate binary head, for the reason given at the top of this page. Final results come from a 5-fold ensemble.

Because the masks are inputs rather than post-hoc explanations, the answer to why did it say severe is a set of segmentations you can look at.

Limitations

Stated plainly, and before a reviewer states them.

  • EyePACS dominates the pooled test set. It is 91% of it, with Stage-A accuracy 0.476. The pooled figure is essentially the EyePACS figure wearing a disguise. The per-dataset table is the primary result; the pooled number is not.
  • EyePACS labels are single-grader with no adjudication. There is no ceiling estimate for label noise on the largest component of the evaluation.
  • Stage I discards 3,161 true DR cases. Whatever Stage III achieves, end-to-end sensitivity is capped at 0.813 by that screening loss.
  • FGADR specificity is not reported. It would be computed on almost no true negatives, so it is N/A rather than flattering.