# BioHCL-SSSM Project Deliverables (v2 - Supervisor Alignment Release) This directory (**`deliverables/v2/`**) represents the finalized, restructured submission addressing all feedback received from Tathagat Sir on Nov 4, 2026. --- ## Directory Index ``` deliverables/v2/ ├── README.md # This navigation guide ├── tathagat_sir_revision_report.md # Master revision report with all formatted markdown tables ├── 1_sota_benchmarks_master.csv # Comparative SOTA Benchmarks (15 Models) ├── 2_table1_component_ablation.csv # Table 1: Macro Component Ablation (SSSM vs Attention vs CNN) ├── 3_table2_layer_hyperparameter_deduction.csv # Table 2: Proposed SSSM Layer Hyperparameter Deduction ├── 4_table3_loss_function_weighting.csv # Table 3: Loss Function Formulations & Class Weighting Dynamics ├── 5_10_fold_cross_validation_metrics.csv # Approved 10-Fold Cross-Validation Metrics & Losses └── figures/ # Publication-grade evaluation figures ├── 10_fold_overlay_roc_curve.png # 10-Fold ROC curve with variance envelopes ├── 10_fold_overlay_pr_curve.png # 10-Fold PR curve with variance envelopes ├── 10_fold_confusion_matrix.png # Pooled 10-Fold Confusion Matrix (5,232 images) ├── 10_fold_global_roc_curve.png # Pooled global ROC ├── 10_fold_global_pr_curve.png # Pooled global PR ├── 10_fold_trajectory_plot.png # Training/validation curves across folds ├── layerwise_tradeoff_plot.png # Latency / FLOPs vs F1 Pareto frontier ├── test_student_biohcl_sssm_cm.png # Official Test Set Confusion Matrix (BioHCL-SSSM) ├── test_student_biohcl_sssm_roc.png # Official Test Set ROC Curve (BioHCL-SSSM) ├── test_baseline_resnet50_cm.png # Official Test Set Confusion Matrix (ResNet-50) ├── test_baseline_resnet50_roc.png # Official Test Set ROC Curve (ResNet-50) ├── test_bioprior_ablation_cm.png # Official Test Set Confusion Matrix (BioPrior Ablated) └── test_roc_pr_comparison.png # Official Test Set ROC/PR Comparison ``` --- ## Executive Summary of Deliverables in v2 ### 1. `1_sota_benchmarks_master.csv` (15-Model SOTA Benchmark) * **Supervisor Directive Addressed:** *"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"* * **Coverage:** Evaluated on the exact 624-image Kermany test split across 15 models: * **Classic CNNs:** Inception-V3, ResNet-18, DenseNet-121, ResNet-50, Xception, GoogLeNet, AlexNet, VGG-16. * **Modern Lightweight CNNs:** MobileNetV3-Large, MobileNet-v1. * **Vision Transformers:** ViT-Base (ViT-B/16), DeiT-Small, Swin-T. * **Proposed Hybrid:** BioHCL-SSSM w/o Bio-Prior, BioHCL-SSSM (Proposed Best). * **Key Takeaway:** BioHCL-SSSM achieves **96.47% test accuracy, 97.22% F1, and 99.37% ROC-AUC** with **14.01M params and 8.56 GFLOPs**, outperforming ViT-Base (86.57M params, 92.47% acc) with **83.8% fewer parameters**, and surpassing Swin-T (27.52M params, 95.99% acc) with **49.1% fewer parameters**. ### 2. `2_table1_component_ablation.csv` (Macro Component Ablation) * **Supervisor Directive Addressed:** *"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. Layer wise is wrong. When we say layer wise it's not layer by layer but components by components"* * **Key Takeaway:** * SSSM vs. Attention (MHSA): Attention costs **+50.0% more compute (12.84 vs 8.56 GFLOPs)** and **+59.8% higher latency (78.4ms vs 49.06ms)**, but drops F1 by **-0.0269** due to $O(N^2)$ quadratic complexity. * SSSM vs. Depthwise Conv: Local conv is 34.6% faster, but drops F1 by **-0.0703** due to lack of bilateral receptive field. * Dual vs. Single Backbone: Dropping DenseNet-121 drops F1 by **-0.2055**. * Bio-Prior: Removing anatomical prior drops F1 by **-0.0775**. ### 3. `3_table2_layer_hyperparameter_deduction.csv` (Proposed SSSM Layer Dynamics) * **Supervisor Directive Addressed:** *"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"* * **Key Takeaway:** * **Block Depth $L$:** Swept $L \in \{1, 2, 3, 4, 6\}$; deduced **$L = 3$** as the optimal Pareto frontier (14.01M params, 49.06ms latency, 0.9987 Val F1). * **State Dimension $d_{state}$:** Swept $d_{state} \in \{16, 32, 64, 128\}$; deduced **$d_{state} = 64$** (0.9987 F1); $d_{state} = 16$ bottlenecks (0.9580 F1), $d_{state} = 128$ yields no gain with +24% latency. * **Expansion Factor $E$:** Deduced **$E = 2$** (standard Mamba expansion); $E=4$ caused parameter explosion (18.25M params) and overfitting. * **Residual Highway:** Additive residual skip connection prevents ODE gradient dissipation (+8.67% F1 over no-skip). ### 4. `4_table3_loss_function_weighting.csv` (Loss Functions & Weighting Dynamics) * **Supervisor Directive Addressed:** *"Then table 3 is loss function which u are using what weighted metrics are taken for granted etc"* * **Key Takeaway:** * **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. * **Weighted CE ($w = [1.95, 0.67]$):** Elevates specificity to **93.16% (+11.54%)**. * **Focal Loss:** Hard-mining suppresses easy cases; focuses on subtle infiltrates (94.07% acc, 0.9529 F1). * **Bio-PINN Loss (Proposed):** Enforces anatomical lung masks, achieving **98.21% Sensitivity, 93.59% Specificity, and 99.37% ROC-AUC**. ### 5. `5_10_fold_cross_validation_metrics.csv` (Approved 10-Fold CV) * **Supervisor Status:** **Approved as correct by Tathagat Sir** (*"CV is correct"*). * **Key Takeaway:** Pooled across 5,232 images: **3,862 TP, 1,337 TN, 12 FP, 21 FN** (**99.37% Global Accuracy**, only 33 errors). Peak fold was **Fold 2 (99.81% accuracy, 0.9987 F1)**.