Leakage-free retrain + evaluation

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+best_finetune.keras filter=lfs diff=lfs merge=lfs -text
+best_head.keras filter=lfs diff=lfs merge=lfs -text
+cataract_final.keras filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,63 @@
+---
+tags: [image-classification, medical-imaging, ophthalmology, cataract]
+library_name: keras
+pipeline_tag: image-classification
+---
+
+# Cataract Severity Classification
+
+EfficientNetB0 with a squeeze-excitation attention block, fine-tuned to grade
+anterior-segment eye images into four severity levels: Normal, Mild, Moderate,
+Severe.
+
+NeuroGSD group project, Faculty of Information Technology, University of Moratuwa.
+
+## Intended use
+
+Research and academic evaluation only. **Not a medical device.** Not clinically
+validated, no regulatory clearance, must not inform decisions about real patients.
+
+## Results (leakage-free held-out test set, n=54)
+
+| Metric | Value | 95% CI |
+|---|---|---|
+| Accuracy | 0.833 | 0.741–0.926 |
+| Balanced accuracy | 0.721 | 0.554–0.862 |
+| Quadratic weighted kappa | 0.697 | 0.321–0.921 |
+| Macro AUC | 0.971 | — |
+| Within-one-grade agreement | 0.981 | — |
+
+## Data and splitting
+
+357 unique images. The source export applied augmentation *before* splitting, so
+several augmented copies of one photograph existed. Splits are therefore grouped
+by original image ID: training keeps all augmented copies, while validation and
+test keep one representative per group. This eliminates train/test leakage that
+would otherwise inflate every metric substantially.
+
+## Input
+
+224x224 RGB, raw 0-255 float (EfficientNet normalisation is inside the graph).
+Preprocessing must match training exactly:
+
+1. Median blur, kernel 3
+2. CLAHE on the LAB L channel, clip 2.0, 8x8 tiles
+3. Resize to 224x224, INTER_AREA
+
+No ROI crop and no circular mask — these images are rectangular anterior-segment
+photographs, and a circular mask removes roughly 29% of real content.
+
+## Limitations
+
+- Only 357 unique images; the Normal grade has just 29, so its metrics rest on a
+  handful of test cases and the confidence intervals are wide.
+- Single-source dataset; other cameras, lighting and populations are unverified.
+- Labels were graded by the student team and reviewed by a domain expert, not
+  independently double-graded.
+- Grade 2 accounts for the majority of unique images; apparent class balance in
+  the raw file counts was an artefact of augmentation.
+
+## Citation
+
+NeuroGSD (2026). AI Based Early Detection System for Common Eye Diseases Using
+Medical Image Analysis. University of Moratuwa.

best_finetune.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a112d0353a7b203caf07ce64079ee15dedbc428e98f459fac21fba659347366
+size 44753233

best_head.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c6885dbdbc65914e048ec28d622e6d081ab234679389188a9ae04ae1a178f30
+size 23539113

cataract_final.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55ef340bc7f3bcb31d84b089ee569fbbba569aeb240f3ee08a906ab2223af4b9
+size 44753233

evaluation.json

@@ -0,0 +1,95 @@
+{
+  "n_test": 54,
+  "tta": true,
+  "leakage_free": true,
+  "accuracy": 0.8333333333333334,
+  "balanced_accuracy": 0.7214285714285713,
+  "qwk": 0.6974291364535268,
+  "mcc": 0.6764422721683907,
+  "macro_auc": 0.9705007198848183,
+  "mae_grades": 0.2037037037037037,
+  "within_one_grade": 0.9814814814814815,
+  "log_loss": 0.579186409720307,
+  "bootstrap_ci": {
+    "accuracy": {
+      "mean": 0.8360277777777778,
+      "lo": 0.7407407407407407,
+      "hi": 0.9259259259259259
+    },
+    "balanced_accuracy": {
+      "mean": 0.7190323405395228,
+      "lo": 0.5544188428059396,
+      "hi": 0.8618421052631579
+    },
+    "qwk": {
+      "mean": 0.6890407941887947,
+      "lo": 0.32109167143522915,
+      "hi": 0.9207563769158413
+    }
+  },
+  "confusion_matrix": [
+    [
+      4,
+      0,
+      0,
+      1
+    ],
+    [
+      0,
+      6,
+      1,
+      0
+    ],
+    [
+      0,
+      2,
+      33,
+      0
+    ],
+    [
+      0,
+      0,
+      5,
+      2
+    ]
+  ],
+  "per_class": {
+    "0": {
+      "precision": 1.0,
+      "recall": 0.8,
+      "f1-score": 0.8888888888888888,
+      "support": 5.0
+    },
+    "1": {
+      "precision": 0.75,
+      "recall": 0.8571428571428571,
+      "f1-score": 0.8,
+      "support": 7.0
+    },
+    "2": {
+      "precision": 0.8461538461538461,
+      "recall": 0.9428571428571428,
+      "f1-score": 0.8918918918918919,
+      "support": 35.0
+    },
+    "3": {
+      "precision": 0.6666666666666666,
+      "recall": 0.2857142857142857,
+      "f1-score": 0.4,
+      "support": 7.0
+    },
+    "accuracy": 0.8333333333333334,
+    "macro avg": {
+      "precision": 0.8157051282051282,
+      "recall": 0.7214285714285713,
+      "f1-score": 0.7451951951951952,
+      "support": 54.0
+    },
+    "weighted avg": {
+      "precision": 0.8246676163342829,
+      "recall": 0.8333333333333334,
+      "f1-score": 0.8159381603826048,
+      "support": 54.0
+    }
+  }
+}

log_finetune.csv

@@ -0,0 +1,12 @@
+epoch,acc,learning_rate,loss,val_acc,val_loss,val_qwk
+0,0.930656909942627,9.999999747378752e-06,0.3946684002876282,0.8888888955116272,0.6446592807769775,0.893491124260355
+1,0.953441321849823,9.999999747378752e-06,0.36877989768981934,0.8888888955116272,0.6242743730545044,0.893491124260355
+2,0.953441321849823,9.999999747378752e-06,0.37556153535842896,0.8703703880310059,0.6032083034515381,0.8771929824561403
+3,0.9331983923912048,9.999999747378752e-06,0.3719872236251831,0.8703703880310059,0.5901629328727722,0.8771929824561403
+4,0.9655870199203491,9.999999747378752e-06,0.35033589601516724,0.8703703880310059,0.5708237290382385,0.8771929824561403
+5,0.9615384340286255,9.999999747378752e-06,0.35351890325546265,0.8703703880310059,0.5533149838447571,0.8771929824561403
+6,0.9676113128662109,9.999999747378752e-06,0.3405035436153412,0.8703703880310059,0.5340921878814697,0.8771929824561403
+7,0.9676113128662109,9.999999747378752e-06,0.34166622161865234,0.8703703880310059,0.5190687775611877,0.8771929824561403
+8,0.9635627269744873,9.999999747378752e-06,0.33416253328323364,0.8703703880310059,0.507893443107605,0.8771929824561403
+9,0.9595141410827637,9.999999747378752e-06,0.352279394865036,0.8518518805503845,0.4947706162929535,0.861271676300578
+10,0.9554656147956848,9.999999747378752e-06,0.345771849155426,0.8888888955116272,0.48168063163757324,0.893491124260355

log_head.csv

@@ -0,0 +1,17 @@
+epoch,acc,learning_rate,loss,val_acc,val_loss,val_qwk
+0,0.7206477522850037,0.0010000000474974513,0.9701550602912903,0.48148149251937866,1.0439261198043823,0.5194805194805194
+1,0.862348198890686,0.0010000000474974513,0.5114316940307617,0.7407407164573669,0.8778523802757263,0.6101694915254237
+2,0.931174099445343,0.0010000000474974513,0.44203224778175354,0.8333333134651184,0.7314485907554626,0.713375796178344
+3,0.9210526347160339,0.0010000000474974513,0.40995514392852783,0.8703703880310059,0.7296186089515686,0.8203592814371257
+4,0.9251012206077576,0.0010000000474974513,0.4605046808719635,0.8888888955116272,0.6689417362213135,0.8867924528301887
+5,0.9473684430122375,0.0010000000474974513,0.391880601644516,0.8888888955116272,0.6693353056907654,0.893491124260355
+6,0.9352226853370667,0.0010000000474974513,0.39837750792503357,0.7592592835426331,0.7415996193885803,0.6815642458100559
+7,0.9331983923912048,0.0010000000474974513,0.4036140739917755,0.7962962985038757,0.6897245049476624,0.718232044198895
+8,0.9615384340286255,0.0010000000474974513,0.3714412450790405,0.7407407164573669,0.7293246388435364,0.679144385026738
+9,0.9696356058120728,0.0010000000474974513,0.34637898206710815,0.7592592835426331,0.6768876314163208,0.6918918918918919
+10,0.9635627269744873,0.0003000000142492354,0.33492517471313477,0.7962962985038757,0.6475585699081421,0.772972972972973
+11,0.9655870199203491,0.0003000000142492354,0.3481481075286865,0.7962962985038757,0.6218593120574951,0.772972972972973
+12,0.9655870199203491,0.0003000000142492354,0.32996681332588196,0.8148148059844971,0.5797955989837646,0.7257142857142858
+13,0.9797570705413818,0.0003000000142492354,0.3153730630874634,0.8333333134651184,0.5787845849990845,0.850828729281768
+14,0.9615384340286255,0.0003000000142492354,0.3239787220954895,0.7777777910232544,0.5722655653953552,0.6983240223463687
+15,0.9676113128662109,0.0003000000142492354,0.31934496760368347,0.8333333134651184,0.5348048806190491,0.7966101694915254

serving_config.json

@@ -0,0 +1,28 @@
+{
+  "class_names": [
+    "0",
+    "1",
+    "2",
+    "3"
+  ],
+  "img_size": 224,
+  "backbone": "efficientnetb0",
+  "attention": "se",
+  "ordinal": true,
+  "preprocess": {
+    "median_blur": 3,
+    "clahe_clip": 2.0,
+    "clahe_grid": [
+      8,
+      8
+    ],
+    "crop_roi": false,
+    "circle_mask": false
+  },
+  "risk_band": {
+    "0": "None",
+    "1": "Low",
+    "2": "Medium",
+    "3": "High"
+  }
+}