    
> A research-oriented two-stage deep learning framework for brain MRI analysis, combining **pixel-level tumor segmentation** with **tumor / non-tumor classification, uncertainty estimation, confidence modeling, and explainable attention**. --- ## 1. Abstract This project implements two complementary neural-network pipelines for brain tumor analysis from 2D MRI slices. The first branch performs **tumor segmentation** using a customized **DeepLabV3+ with ResNet101V2**, augmented with **Atrous Spatial Pyramid Pooling (ASPP)**, **CBAM attention**, a lightweight separable-convolution decoder, a hybrid loss, and morphological post-processing. The second branch performs **tumor classification** using a pretrained **EfficientNetV2-B3-derived backbone** with two-phase fine-tuning, tumor-centric sampling, strong Albumentations augmentation, evidential uncertainty estimation, explainable self-attention, and a dedicated confidence head. Model selection is performed using **Quadratic Weighted Kappa (QWK)** rather than accuracy alone. The overall design separates three complementary objectives: ```text Brain MRI Slice │ ├──────────────► Localization │ DeepLabV3+ │ ResNet101V2 │ CBAM + ASPP │ └──────────────► Recognition EfficientNetV2 Self-Attention Evidential Head Confidence Head ``` The result is a research framework that combines **localization, recognition, reliability estimation, and interpretability** rather than treating brain tumor analysis as a single classification problem. --- # 2. Research Motivation Brain tumor analysis from MRI is challenging because tumor regions can show strong variation in size, shape, intensity, and visual appearance. In addition, the foreground tumor region is often much smaller than the background, creating substantial class imbalance for segmentation. A single classification probability is also insufficient for a reliability-aware medical imaging system. For that reason, the project separates the problem into two questions: > **Where is the tumor?** and > **Does this slice contain a tumor, and how reliable is that prediction?** This leads to the following conceptual structure: ```text Brain MRI │ ┌─────────┴─────────┐ ▼ ▼ Segmentation Classification │ │ ▼ ▼ Tumor Mask Class Probability │ ┌─────────────┼─────────────┐ ▼ ▼ ▼ Uncertainty Confidence Attention ``` --- # 3. Main Contributions ## 3.1 Segmentation Branch The segmentation model combines: - DeepLabV3+ - ResNet101V2 ImageNet backbone - ASPP with dilation rates 6, 12, and 18 - Image-level pooling branch - CBAM attention on semantic and low-level features - Separable convolution decoder - Dynamic upsampling / resize alignment - Hybrid Focal-Tversky + Generalized Dice + Boundary + Focal loss - Morphological post-processing - Region, boundary, and detection-oriented metrics ## 3.2 Classification Branch The classification model combines: - Pretrained EfficientNetV2-B3-derived representation - Mixed-precision training - Two-phase freeze / unfreeze fine-tuning - Tumor-centric batch sampling - Class-weighted learning - Strong Albumentations augmentation - Explainable spatial self-attention - Evidential probability modeling - Epistemic uncertainty - Aleatoric uncertainty - Confidence estimation - QWK-based validation and checkpoint selection --- # 4. Overall Architecture ```text ┌────────────────────────────────────────────────────────────────────┐ │ BRAIN MRI SLICE │ └───────────────────────────────┬────────────────────────────────────┘ │ ┌──────────────────┴──────────────────┐ │ │ ▼ ▼ ┌─────────────────────────┐ ┌──────────────────────────────┐ │ SEGMENTATION BRANCH │ │ CLASSIFICATION BRANCH │ │ │ │ │ │ 1ch → 3ch replication │ │ EfficientNetV2 representation│ │ │ │ │ │ │ │ ResNet101V2 │ │ ▼ │ │ │ │ │ Explainable Self-Attention │ │ ┌─────┴─────┐ │ │ │ │ │ ▼ ▼ │ │ Global Features │ │ Low-level High-level │ │ │ │ │ │ │ │ │ ┌──────┴──────┐ │ │ CBAM ASPP │ │ ▼ ▼ │ │ │ │ │ │ Classification Evidential │ │ │ CBAM │ │ │ │ │ │ └──────┬────┘ │ │ └──────┬──────┘ │ │ ▼ │ │ ▼ │ │ Feature Fusion │ │ Confidence Head │ │ │ │ │ │ │ Decoder │ └──────────────────────────────┘ │ │ │ │ ▼ │ │ Tumor Mask │ └─────────────────────────┘ ``` --- # 5. Segmentation Model ## 5.1 Input Strategy The segmentation pipeline keeps the dataset representation as **one grayscale channel**: ```text Input = 256 × 256 × 1 ``` Because ResNet101V2 is loaded with ImageNet weights, the input is replicated to three channels **inside the model**: ```text Gray MRI │ ├─────┐ ├─────┼──► Concatenate └─────┘ │ ▼ 256 × 256 × 3 │ ▼ ResNet101V2 ``` This avoids changing the pretrained backbone while preserving a compact single-channel data representation at the input. --- ## 5.2 Encoder The backbone is: ```python ResNet101V2( include_top=False, weights="imagenet" ) ``` Two feature levels are extracted: | Feature | Layer | Role | |---|---|---| | Low-level | `conv2_block3_out` | Spatial detail / boundaries | | High-level | `base_model.output` | Semantic context | Batch Normalization layers in the backbone are kept non-trainable, while other backbone layers remain trainable. --- # 6. ASPP Multi-Scale Context The high-level representation enters an ASPP module containing parallel branches: ```text High-level Feature Map │ ┌────────────────────┼────────────────────┐ │ │ │ │ │ ▼ ▼ ▼ ▼ ▼ 1×1 d=6 d=12 d=18 Image Pool │ │ │ │ │ └──────────┴─────────┴─────────┴──────────┘ │ ▼ Concatenate │ ▼ 1×1 Conv │ ▼ Dropout ``` The use of multiple dilation rates allows the network to capture tumor context at different effective receptive fields. --- # 7. CBAM Attention CBAM is applied to both the ASPP output and the low-level feature path. ```text Input Feature │ ▼ Channel Attention │ ▼ Spatial Attention │ ▼ Refined Feature ``` This provides two complementary forms of feature selection: - **Channel attention**: which feature channels matter - **Spatial attention**: where relevant structures are located The intention is to improve feature selectivity before decoder fusion. --- # 8. Decoder and Shape Alignment A common engineering issue in encoder-decoder segmentation networks is spatial mismatch between high-level and low-level feature maps. This implementation explicitly computes the upsampling factor from the actual tensor dimensions and, when necessary, performs a final resize before concatenation. ```text High-level / ASPP │ ▼ Dynamic Upsampling │ ▼ Shape Check │ ┌────┴────┐ │ │ Match Mismatch │ │ │ ▼ │ tf.image.resize │ │ └────┬────┘ ▼ Concatenate with low-level features ``` This is a practical improvement that makes the architecture less brittle when tensor dimensions change. --- # 9. Segmentation Objective The main segmentation loss is a weighted hybrid: ```text Lseg = 0.50 LFT + 0.35 LGD + 0.10 LBoundary + 0.05 LFocal ``` where: - `LFT` = Focal Tversky loss - `LGD` = Generalized Dice loss - `LBoundary` = Sobel boundary loss - `LFocal` = Focal loss ### Focal Tversky Uses: ```text alpha = 0.8 beta = 0.2 gamma = 0.75 ``` to control the relative penalty for false negatives and false positives. ### Generalized Dice Improves robustness to foreground/background imbalance. ### Boundary Loss Compares Sobel-derived edge maps of the ground-truth and predicted masks. ### Focal Loss Adds additional emphasis to hard pixels. The overall design therefore optimizes both **region agreement** and **boundary fidelity**. --- # 10. Segmentation Training Configuration | Parameter | Value | |---|---:| | Resolution | `256 × 256` | | Batch size | `16` | | Epochs | `60` | | Initial LR | `1e-4` | | Weight decay | `5e-4` | | Gradient clipping | `clipnorm=1.0` | | LR schedule | CosineDecayRestarts | | Dropout | `0.7 / 0.6 / 0.5` | | Training subset | `30%` | | Validation split | `25%` temporary split, then half for validation/test | The relatively small batch size is chosen to reduce memory pressure caused by the ResNet101V2-based architecture. --- # 11. Segmentation Post-Processing The raw probability map is post-processed as follows: ```text Sigmoid Probability Map │ ▼ Threshold = 0.25 │ ▼ Morphological Closing │ ▼ Connected Components │ ▼ Remove components < 100 px │ ▼ Light Dilation │ ▼ Final Mask ``` This stage is designed to remove isolated noise and improve mask continuity. --- # 12. Segmentation Metrics The evaluation includes: | Metric | Interpretation | |---|---| | Dice | Overlap between prediction and ground truth | | Generalized Dice | Imbalance-aware overlap | | Weighted Dice | Foreground-prioritized overlap | | IoU | Jaccard similarity | | Boundary IoU | Boundary agreement | | Sensitivity | Tumor recall | | Specificity | Background rejection | | Precision | False-positive control | | F1 | Precision-recall balance | The pipeline stores per-slice results in: ```text test_results_optimized.csv ``` and generates qualitative examples containing: ```text Input | Ground Truth | Prediction | Overlay ``` --- # 13. Classification Model The classifier uses a previously trained EfficientNetV2-based model as a feature extractor and builds a new multi-output prediction head. The pipeline receives: ```text 299 × 299 × 3 ``` The three channels are currently constructed by repeating the same MRI slice because true adjacent-slice information is not available in the generator. ```text Central slice │ ├──────── Channel 1 ├──────── Channel 2 └──────── Channel 3 ``` Therefore, this is a **2D / pseudo-3-channel representation**, not true 3D context. --- # 14. Transfer Learning and Fine-Tuning The pretrained model is loaded with custom objects including: - `AttentionVisualizer` - `MCDropout` - `EvidentialLoss` - `EvidentialLayer` - `ExplainableSelfAttention` - `MaxProbLayer` The original feature representation is reused while a new classification / uncertainty head is attached. This creates a transfer-learning pipeline: ```text Pretrained EfficientNetV2 │ ▼ Feature Extraction │ ▼ New Task-Specific Head │ ┌──────┼─────────┐ ▼ ▼ ▼ Class Evidence Confidence ``` --- # 15. Two-Phase Classification Training ## Phase 1 — Head Adaptation The pretrained backbone is frozen and only the new head is optimized. ```text Backbone = Frozen Head = Trainable LR = 2e-5 Epochs = 10 ``` ## Phase 2 — Full Fine-Tuning The network is then unfrozen: ```text Backbone = Trainable Head = Trainable LR = 5e-6 ``` The lower learning rate reduces the risk of destroying useful pretrained representations. --- # 16. Data Strategy for Classification The classification data pipeline includes several improvements. ### Volume-aware sampling Each volume contributes at least one slice before the remaining samples are selected. ### Stratified sampling Additional data are sampled according to the target class. ### Tumor-centric batches Training batches target: ```text 60% tumor 40% non-tumor ``` ### Class weighting The generator additionally computes class weights from the selected training set. ### Augmentation Albumentations introduces geometric, intensity, blur, noise, and elastic variations. --- # 17. Explainable Self-Attention The classification model includes a custom spatial self-attention layer. Given a feature map: ```text H × W × C ``` it is flattened to spatial tokens: ```text (H × W) × C ``` and transformed into query, key, and value representations. The attention mechanism is: ```text Q = Wq X K = Wk X V = Wv X A = softmax(QKᵀ / √d) Y = AV ``` The output is reshaped back into image space, and the attention layer can return a normalized spatial attention map. This provides an explicit mechanism for visualizing where the representation is attending. --- # 18. Evidential Uncertainty The classifier does not stop at softmax probabilities. The evidential head computes positive evidence using a softplus activation: ```text Feature Vector │ ▼ Dense Layer │ ▼ Softplus │ ▼ Evidence │ ▼ alpha = evidence + 1 │ ▼ Normalized Probability ``` Two uncertainty signals are exposed: ### Epistemic Uncertainty A signal related to the model's lack of knowledge. ### Aleatoric Uncertainty A signal related to uncertainty inherent in the observation. These signals are then used together with maximum class probability to construct the confidence branch. --- # 19. Confidence Branch The confidence head uses: ```text max(class probability) │ ├──────────────┐ ▼ ▼ epistemic uncertainty aleatoric uncertainty │ │ └──────┬───────┘ ▼ Dense Network │ ▼ Confidence ∈ [0,1] ``` This is intended to provide a dedicated model confidence signal rather than interpreting softmax probability alone as certainty. --- # 20. Multi-Task Classification Objective The compiled model optimizes three supervised outputs: ```text Lcls = 2.00 LCE + 0.05 LEvidential + 0.01 LConfidence ``` where: - `LCE` is categorical cross-entropy with label smoothing `0.05` - `LEvidential` is the custom evidential loss - `LConfidence` is mean-squared error The classification objective remains dominant, while the evidential and confidence branches act as auxiliary learning signals. --- # 21. QWK-Centered Model Selection The custom `ImprovedQWKEvaluation` callback evaluates validation predictions after every epoch. It computes: ```text Quadratic Weighted Kappa Accuracy Weighted F1 Per-class F1 Per-class Recall ``` The checkpoint criterion is: ```text Best Validation QWK ``` Training uses: ```text EarlyStopping(patience=10) ReduceLROnPlateau(patience=5) ``` This makes the optimization process more aligned with the selected validation objective than monitoring accuracy alone. --- # 22. Classification Outputs The final model exposes multiple outputs: | Output | Meaning | |---|---| | `classification` | Standard 2-class softmax probabilities | | `evidential_prob` | Evidential class probabilities | | `epistemic_uncertainty` | Epistemic uncertainty estimate | | `aleatoric_uncertainty` | Aleatoric uncertainty estimate | | `confidence_score` | Dedicated confidence prediction | | `attention_map` | Spatial explanation map | | `conv_features` | Intermediate feature representation | This makes the model suitable for both prediction and downstream analysis. --- # 23. Outputs and Artifacts ## Segmentation ```text brain_tumor_models/ ├── best_deeplabv3plus_resnet101v2_model.keras ├── final_optimized_model.keras ├── training_log.csv ├── test_results_optimized.csv ├── training_history_optimized.png └── visualizations/ ├── test_sample_0.png ├── test_sample_1.png └── ... ``` ## Classification ```text brain_tumor_models/ ├── phase1_finetuned_best.keras ├── phase2_finetuned_best.keras ├── final_finetuned_model.keras ├── phase1_finetuned_training.csv ├── phase2_finetuned_training.csv ├── test_confusion_matrix.png ├── test_classification_report.csv ├── confidence_analysis.png └── training_history.png ``` --- # 24. Experimental Configuration Summary ## Segmentation | Component | Configuration | |---|---| | Resolution | `256 × 256` | | Input channels | `1` | | Backbone | ResNet101V2 | | Architecture | DeepLabV3+ | | Attention | CBAM | | ASPP dilation | `6 / 12 / 18` | | Batch size | `16` | | Epochs | `60` | | Initial LR | `1e-4` | | Weight decay | `5e-4` | | LR scheduler | CosineDecayRestarts | | Dataset sample | `30%` | | Output | Binary tumor mask | ## Classification | Component | Configuration | |---|---| | Resolution | `299 × 299` | | Input channels | `3` (repeated slice) | | Backbone | EfficientNetV2-B3-derived pretrained model | | Classes | `2` | | Batch size | `10` | | Dataset sample | `10%` | | Tumor-centric ratio | `60%` | | Precision | Mixed FP16 | | Phase 1 LR | `2e-5` | | Phase 2 LR | `5e-6` | | Total epochs | `30` | | Model selection | QWK | --- # 25. Methodological Advantages The combined design has several strengths. ### Multi-scale segmentation ASPP captures spatial context at multiple dilation rates, while the decoder restores fine details through low-level feature fusion. ### Attention-guided representation CBAM helps refine both semantic and spatial feature representations. ### Imbalance-aware optimization The segmentation loss explicitly considers overlap, boundary quality, and hard examples; the classifier uses balanced sampling and class weighting. ### Transfer learning The classifier reuses a pretrained EfficientNetV2 representation rather than learning all features from scratch. ### Reliability-aware classification The evidential branch, uncertainty signals, and confidence head provide information beyond the class label itself. ### Explainability The self-attention component exposes a spatial attention representation suitable for visualization and qualitative analysis. --- # 26. Important Limitations This repository should be considered a **research and experimental framework**, not a clinically validated diagnostic system. ### Slice-level processing Both pipelines primarily operate on 2D slices. ### Pseudo-3-channel classification input The classifier currently repeats the same slice three times: ```text Channel 1 = central slice Channel 2 = central slice Channel 3 = central slice ``` Therefore, it does not yet model true inter-slice anatomical context. ### Sampling subset The training scripts use sampled portions of the dataset (`30%` for segmentation and `10%` for classification). Performance should therefore be re-evaluated when using the complete dataset. ### Leakage considerations For research-grade evaluation, splitting should ideally be performed **at the patient / volume level**, not only at the slice level, to avoid highly correlated slices from the same volume appearing across train, validation, and test sets. ### Clinical validation No clinical diagnostic claim should be inferred without external validation, calibration analysis, multi-center evaluation, and appropriate clinical study design. --- # 27. Future Work A natural extension is to move from slice-level to volume-aware analysis: ```text 2D Slice │ ▼ 2.5D Adjacent Slices │ ▼ 3D Volume Modeling │ ▼ Volume-level Classification │ ▼ Integrated Segmentation + Classification ``` Potential research directions include: - True adjacent-slice 2.5D input - 3D segmentation architectures - Patient / volume-level splitting - External-dataset validation - Probability calibration - Uncertainty calibration - Attention faithfulness evaluation - Joint segmentation-classification learning - Volume-level aggregation - Ensemble-based uncertainty estimation --- # 28. Reproducibility Install the main dependencies: ```bash pip install tensorflow numpy pandas scipy h5py opencv-python scikit-learn matplotlib seaborn albumentations ``` Expected CSV columns: ```text slice_path target ``` Expected HDF5 content for segmentation: ```text image mask ``` Expected HDF5 content for classification: ```text image ``` Configure the dataset and pretrained-model paths in the scripts before execution. --- # 29. Research Summary The project can be summarized as a four-layer analytical framework: ```text ┌──────────────────────────────────────────────┐ │ Brain MRI Input │ └──────────────────────┬───────────────────────┘ │ ┌────────────┴────────────┐ ▼ ▼ ┌────────────────────┐ ┌──────────────────────┐ │ Tumor Localization │ │ Tumor Classification │ │ DeepLabV3+ │ │ EfficientNetV2 │ │ ResNet101V2 │ │ Self-Attention │ │ ASPP + CBAM │ │ Evidential Learning │ └──────────┬─────────┘ └───────────┬──────────┘ │ │ ▼ ▼ Tumor Mask Class + Uncertainty + Confidence + Attention ``` The central research idea is therefore: > **Localization + Recognition + Uncertainty + Explainability** rather than a single black-box classification output. --- ## Citation ```bibtex @software{brain_tumor_ai_pipeline_2026, title = {Brain Tumor AI Pipeline: DeepLabV3+ Segmentation and Explainable EfficientNetV2 Classification}, year = {2026}, note = {Research-oriented brain MRI analysis framework} } ``` ---**Brain Tumor AI Pipeline · 2026** *Segmentation · Classification · Uncertainty · Explainable AI*