Efficient Deep Learning For Brain Tumour Segmentation

Transcript

1. Efficient Deep Learning for Brain Tumor Segmentation Machine Learning Lagos Olamide Lawal Machine Learning Engineer & Independent Researcher @ML Collective https://em07adoz.vercel.app AUGUST 2026 1
2. AUGUST 2025 What We'll Explore EFFICIENT DEEP LEARNING FOR BRAIN TUMOR SEGMENTATION 01 — The Problem 02 — The Baseline Brain tumor segmentation & Understanding 3D U-Net and why 3D makes it challenging its computational cost 03 — Making It Efficient 04 — The Trade-offs Architectural strategies for Accuracy, computation, lightweight segmentation memory & inference 05 — Beyond Accuracy 06 — Open Questions What makes a model useful, Where efficient medical AI can not just accurate go next 2
3. 06 THE PROBLEM How accurate is your model? Dice, IoU, Accuracy But how much does that accuracy cost? Computation, Memory, Time, Storage 3
4. Machine Learning “The purpose of computing is insight, not numbers.” Lagos — Richard Hamming What if we stopped treating model performance as a single number? Machine Learning Lagos 4
5. Machine Learning Lagos THE PROBLEM 5
6. THE PROBLEM 00 Brain Tumor Segmentation Multi-modal MRI Deep Learning Model Voxel-level Segmentation The goal is not simply to determine whether a tumor exists, but to identify the exact voxels belonging to tumor regions. 6
7. THE PROBLEM 00 CLASSIFICATION VS DETECTION VS SEGMENTATION Classification “Is there a tumor?” Detection “Where is the tumor?” Segmentation “Which voxels belong to the tumor?” Segmentation requires spatial precision. 7
8. THE PROBLEM 00 WHY IS IT HARD? 01 — 3D Data Large volumetric inputs 02 — Small Structures Tumor regions can occupy a small fraction of the volume 03 — Complex Boundaries Irregular tumor morphology 04 — Multiple Modalities Different MRI sequences reveal different characteristics Better spatial understanding comes with greater computational cost. 8
9. Machine Learning Lagos UNDERSTANDING THE COMPUTATION 9
10. UNDERSTANDING THE COMPUTATION 00 From 2D Images to 3D Volumes 2D 3D H×W D×H×W 155 x 240 x 240 240 x 240 Adding depth doesn't just add information. It adds computation. 10
11. UNDERSTANDING THE COMPUTATION 00 WHAT DOES THE MODEL SEE? 4-Channel 3D Volume the network receives tensors, not a single conventional “image.” 11
12. UNDERSTANDING THE COMPUTATION 00 WHY 3D? 2D 3D Slice-by-slice Volume-level processing ↓ Limited cross-slice context ↓ Captures relationships across slices Spatial context has a computational price. 12
13. UNDERSTANDING THE COMPUTATION 00 OUR BASELINE: 3D U-NET U-Net is a natural baseline because segmentation requires both semantic understanding and spatial precision. 13
14. UNDERSTANDING THE COMPUTATION 00 Encoder The encoder learns increasingly abstract representations. 14
15. UNDERSTANDING THE COMPUTATION 00 Decoder The decoder reconstructs spatial detail needed for precise segmentation. 15
16. UNDERSTANDING THE COMPUTATION 00 WHERE DOES THE COST COME FROM? 3D Convolutions + Large Feature Maps + Many Channels + Multiple Encoder/Decoder Stages + Intermediate Activations High Computational Cost 16
17. UNDERSTANDING THE COMPUTATION 00 PARAMETERS ≠ FLOPs ≠ MEMORY PARAMETERS FLOPs MEMORY What the model learns How much computation it performs What must be stored while it runs Three different questions.Three different measurements. 17
18. UNDERSTANDING THE COMPUTATION 00 THE HIDDEN COST: ACTIVATION MEMORY During training, intermediate activations often need to be retained for backpropagation. Large volume × many channels = large activation footprint 18
19. Machine Learning Lagos MAKING THE MODEL EFFICIENT 19
20. MAKING THE MODEL EFFICIENT 00 WHAT DOES “LIGHTWEIGHT” MEAN? Lightweight ≠ Just Fewer Parameters ↓ Parameters ↓ FLOPs ↓ Memory ↓ Model Size ↓ Latency ↓ Energy Lightweight in what sense? 20
21. MAKING THE MODEL EFFICIENT 00 FOUR WAYS TO SHRINK A MODEL 01. Reduce Width 02. Reduce Depth 03. Use Efficient Operators 04. Process Smaller Inputs Different strategies reduce different costs. 21
22. MAKING THE MODEL EFFICIENT 00 TRAINING EFFICIENCY VS MODEL EFFICIENCY TRAINING DEPLOYMENT Mixed precision Gradient accumulation Gradient checkpointing Smaller batches Fewer parameters Lower FLOPs Quantization Faster inference Making training easier does not necessarily make the final model cheaper. 22
23. Machine Learning Lagos DOES LIGHTWEIGHT MEAN WORSE? 23
24. DOES LIGHTWEIGHT MEAN WORSE? 00 Does Lightweight Mean Worse? Model A Model B Higher Dice Higher memory Higher latency Slightly lower Dice Much lower memory Much faster Which model is better? 24
25. DOES LIGHTWEIGHT MEAN WORSE? 00 Accuracy Needs More Than One Number DICE IOU HD95 25
26. DOES LIGHTWEIGHT MEAN WORSE? 00 Don't Stop at Dice EFFICIENCY SEGMENTATION Dice IoU HD95 MODEL Parameters FLOPs Memory Model Size Latency Evaluation is multi-dimensional. 26
27. Machine “Remember that all models are Learning Lagos wrong; the practical question is how wrong do they have to be to not be useful.” — George E. P. Box Does a 2% performance drop matter if computation falls dramatically? Machine Learning Lagos 27
28. Machine Learning Lagos BEYOND ACCURACY 28
29. BEYOUND ACCURACY 00 We shouldn't only ask: How accurate is the model? We should also ask: How expensive is that accuracy? Is the additional performance worth the additional computation? 29
30. 00 01 3D segmentation is powerful and computationally expensive. FINAL TAKEAWAYS 02 Lightweight design is a series of trade-offs. 03 The best model isn't necessarily the most accurate model. 30
31. 00 Which architectural changes provide the largest efficiency gains? How much segmentation performance can we afford to lose? What We Still Don't Know Does reducing parameters actually reduce inference time? Does lower FLOPs translate to lower memory? How do these trade-offs change across hardware? How well do lightweight models generalize? 31
32. 00 How much accuracy do we actually need and what are we willing to spend to get it? 32
33. Thank You Questions? Machine Learning Lagos AUGUST 2026 33