unknownBrain's picture
Complete medical imaging workspace dump (weights, deliverables, papers, notebooks, code, figures)
eded815 verified
|
Raw History Blame Contribute Delete
7.92 kB

BioHCL-SSSM Project Deliverables Package (Revised per Supervisor Feedback)

This repository package contains the finalized, publication-grade benchmarks, segregated ablation studies, and 10-fold cross-validation results for the BioHCL-SSSM model on the Guangzhou Pediatric Chest X-Ray dataset (Kermany et al., Cell 2018; 5,856 images, 624 official held-out test split).


Deliverables Architecture & Supervisor Alignment

deliverables/
β”œβ”€β”€ README.md                                    # Deliverables index & executive summary
β”œβ”€β”€ tathagat_sir_revision_report.md              # Full report mapped directly to Tathagat Sir's feedback
β”œβ”€β”€ 1_sota_benchmarks_master.csv                 # 15-Model SOTA Benchmark (Literature baselines + Ours)
β”œβ”€β”€ 2_table1_component_ablation.csv              # Table 1: Macro Component Ablation (SSSM vs Attention vs CNN)
β”œβ”€β”€ 3_table2_layer_hyperparameter_deduction.csv  # Table 2: Proposed SSSM Layer Hyperparameter Deduction
β”œβ”€β”€ 4_table3_loss_function_weighting.csv         # Table 3: Loss Function Formulations & Class Weighting Dynamics
β”œβ”€β”€ 5_10_fold_cross_validation_metrics.csv       # Approved 10-Fold CV metrics (Folds 1 to 10 + Mean/Std)
β”œβ”€β”€ raw_sweeps/                                  # Unaggregated raw grid sweep files
β”‚   β”œβ”€β”€ ablation_grid_sweep_64_runs.csv          # 64-run ablation search
β”‚   └── benchmark_profiling_raw.csv              # Latency & FLOPs profiling log
└── figures/                                     # Publication-grade evaluation figures
    β”œβ”€β”€ 10_fold_overlay_roc_curve.png            # 10-Fold ROC with variance envelopes
    β”œβ”€β”€ 10_fold_overlay_pr_curve.png             # 10-Fold PR with variance envelopes
    β”œβ”€β”€ 10_fold_confusion_matrix.png             # Pooled 10-Fold Confusion Matrix (5,232 images)
    β”œβ”€β”€ 10_fold_global_roc_curve.png             # Pooled global ROC
    β”œβ”€β”€ 10_fold_global_pr_curve.png              # Pooled global PR
    β”œβ”€β”€ 10_fold_trajectory_plot.png              # Training/validation curves across folds
    β”œβ”€β”€ layerwise_tradeoff_plot.png              # Latency / FLOPs vs F1 Pareto frontier
    β”œβ”€β”€ test_student_biohcl_sssm_cm.png          # Official Test Set Confusion Matrix (BioHCL-SSSM)
    β”œβ”€β”€ test_student_biohcl_sssm_roc.png         # Official Test Set ROC Curve (BioHCL-SSSM)
    β”œβ”€β”€ test_baseline_resnet50_cm.png            # Official Test Set Confusion Matrix (ResNet-50)
    β”œβ”€β”€ test_baseline_resnet50_roc.png           # Official Test Set ROC Curve (ResNet-50)
    β”œβ”€β”€ test_bioprior_ablation_cm.png            # Official Test Set Confusion Matrix (BioPrior Ablated)
    └── test_roc_pr_comparison.png               # Official Test Set ROC/PR Comparison

1. State-of-the-Art Comparative Benchmarks (1_sota_benchmarks_master.csv)

Directly addresses feedback: "Benchmark - wrong. I need benchmark as in state of the art techniques accuracy as a vanilla model Like Resnet, vit, deit etc around 10 to 15 latest models and ofc ours best model in the sheet"

  • 15 Distinct Models benchmarked on the official 624-image Kermany test split (234 Normal, 390 Pneumonia).
  • Covered Families: Classic CNNs (Inception-V3, ResNet-18, DenseNet-121, ResNet-50, Xception, GoogLeNet, AlexNet, VGG-16), Modern Lightweight CNNs (MobileNetV3, MobileNet-v1), Vision Transformers (ViT-Base, DeiT-Small, Swin-T), and Proposed State-Space Hybrids (BioHCL-SSSM).
  • Key Finding: BioHCL-SSSM achieves 96.47% accuracy, 97.22% F1, and 99.37% ROC-AUC with only 14.01M parameters and 8.56 GFLOPs, outperforming large transformers like ViT-Base (86.57M params, 92.47% acc) and matching/exceeding Swin-T (27.52M params, 95.99% acc) at half the parameter footprint.

2. Table 1: Component-Wise Architectural Ablation (2_table1_component_ablation.csv)

Directly addresses feedback: "Table 1 - for component so ur layers comes there and other alternative like if u don't use SSSM and use attention how much computation is gained how much accuracy is gained or lost etc"

  • SSSM vs. Multi-Head Self-Attention (MHSA): Replacing SSSM with MHSA increases compute by +50.0% FLOPs (12.84 vs 8.56 GFLOPs) and latency by +59.8% (78.4ms vs 49.06ms), while dropping validation F1 by -0.0269 due to quadratic $O(N^2)$ memory bottlenecks and lack of directional scanline continuity.
  • SSSM vs. Depthwise Separable Conv: Replacing SSSM with local convolutions saves compute (-16.8% FLOPs, -16.96ms faster), but suffers an acute -0.0703 F1 drop due to inability to model bilateral consolidations across hemithoraces.
  • Dual-Chain vs. Single-Chain Backbone: Dropping DenseNet-121 drops F1 by -0.2055 (severe underfitting); dropping Hierarchical CNN drops F1 by -0.0483.
  • Clinical Prior Ablation: Disabling anatomical prior drops F1 by -0.0775 due to false-positive over-indexing on peripheral rib/diaphragmatic edges.
  • Fusion Mechanism: Bilinear gated cross-attention outperforms naive additive fusion by +0.0467 F1.

3. Table 2: Proposed SSSM Layer Hyperparameter Deduction (3_table2_layer_hyperparameter_deduction.csv)

Directly addresses feedback: "Table 2 - of ur proposed layer how u can deduce hyperparameters on this. I take only one example which u have used is SSSM blocks or the parameter for ssssm similarly u can have other parameters for other layers. If not atleast table 2"

  • Block Depth ($L$): Swept across $L \in {1, 2, 3, 4, 6}$. Peak Pareto efficiency occurs at $L = 3$ (14.01M params, 49.06ms latency, 0.9987 Val F1). Depths $L=4$ and $L=6$ induce parameter redundancy (+13% to +40%) without accuracy gains.
  • State Dimension ($d_{state}$): Evaluated at $d_{state} \in {16, 32, 64, 128}$. $d_{state} = 16$ creates an information bottleneck (0.9580 F1); $d_{state} = 64$ hits optimal capacity (0.9987 F1); $d_{state} = 128$ increases latency to 61.2ms without performance gains.
  • Channel Expansion Factor ($E$): Evaluated at $E \in {1, 2, 4}$. $E = 2$ offers the optimal continuous ODE transformation; $E=4$ triggers severe over-parameterization (+30% params) prone to overfitting.
  • Residual Connection Strategy: Direct sequential cascading without skips drops F1 to 0.9120 (-8.67%); Additive Identity Highways guarantee stable continuous-to-discrete gradient propagation.

4. Table 3: Loss Function & Class Weighting Dynamics (4_table3_loss_function_weighting.csv)

Directly addresses feedback: "Then table 3 is loss function which u are using what weighted metrics are taken for granted etc"

  • Unweighted CE: Suffers from severe false positive bias (18.38% false alarm rate in healthy children, specificity 81.62%) because it over-indexes on the 72% pneumonia majority class.
  • Class-Weighted CE ($w = [1.95, 0.67]$): Heavily penalizes normal-class misclassifications, elevating specificity to 93.16% (+11.54%).
  • Focal Loss ($\gamma=2.0, \alpha=0.25$): Down-weights easily classified clear lungs, concentrating gradient updates on borderline consolidations (94.07% acc, 0.9529 F1).
  • Bio-PINN Inductive Penalty Loss (Proposed): Integrates anatomical regularization masks, achieving the optimal clinical Pareto frontier: 98.21% Sensitivity, 93.59% Specificity, and 99.37% ROC-AUC.

5. 10-Fold Cross-Validation (5_10_fold_cross_validation_metrics.csv)

Directly addresses feedback: "CV is correct" (Approved by Tathagat Sir)

  • Exact Integer Confusion Matrix: Pooled across 5,232 images: 3,862 TP, 1,337 TN, 12 FP, 21 FN (99.37% global accuracy, only 33 errors).
  • Mean Performance: Accuracy: 99.37% Β± 0.33%, Precision: 0.9969 Β± 0.0036, Recall: 0.9946 Β± 0.0059, Specificity: 0.9911 Β± 0.0103, F1-Score: 0.9957 Β± 0.0022, AUC: 0.9929 Β± 0.0039.