💻 computer science

Quantifying Data Efficiency in YOLO-Based Femur Segmentation for Axial T1-Weighted MRI

This study demonstrates that among YOLO-family models for femur segmentation in T1-weighted MRI, YOLOv11m-seg provides the optimal balance of data efficiency, accuracy, and speed in data-scarce regimes, achieving high performance with as few as 100 annotated slices while showing that architectural differences become negligible with larger datasets.

Original authors: Patrick Gerard Sujeeth, I R Praveen Joe

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Patrick Gerard Sujeeth, I R Praveen Joe

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Medical imaging has long relied on the human eye to trace the boundaries of bones and organs, a task that is both vital for planning surgeries and incredibly tedious for doctors. When a patient needs a hip replacement or a doctor needs to study the health of a femur, they must look at magnetic resonance images, or MRI, and manually outline the bone on every single slice. This process is slow, expensive, and prone to human error, creating a bottleneck that slows down research and treatment. In recent years, computers have learned to do this outlining automatically using deep learning, a type of artificial intelligence that recognizes patterns in images. However, these smart systems usually require thousands of labeled examples to learn, a resource that most hospitals simply do not have. The question facing researchers is whether these systems can be taught to work well with very few examples, and if so, which specific type of computer architecture is best suited for the job when data is scarce.

A team of researchers at the Vellore Institute of Technology in India set out to answer this question by testing how different versions of a popular computer vision system called YOLO perform when asked to segment, or outline, the femur bone in MRI scans. They focused on a specific challenge: the femur is a curved, complex shape, and the MRI images they used were two-dimensional slices where the bone often blends into the surrounding soft tissue. To study this, they started with a very small collection of forty real MRI slices taken from a single healthy volunteer. Since they could not gather thousands of new scans, they used a clever technique to create visual variety from these forty images. They applied a series of safe, anatomically correct changes, such as rotating the images, flipping them, and adjusting the brightness and contrast, to generate hundreds of new training examples. This allowed them to train three different versions of the YOLO system—one older model and two newer, more advanced ones—on datasets ranging from just forty images up to eight hundred.

The researchers discovered that the choice of computer architecture mattered immensely when the amount of training data was low. When they tested the systems with only forty images, the older model performed the most reliably, successfully outlining the bone in most cases. The two newer, more complex models struggled significantly; one of them essentially failed to recognize the bone at all, while the other produced outlines that were barely better than random guesses. This counterintuitive result suggests that more advanced systems, while powerful, can be too sensitive when they do not have enough examples to learn from, causing them to overfit or get confused by the limited data.

The situation changed dramatically when the researchers increased the training data to one hundred images. At this point, the older model collapsed, producing outlines that were so poor they were effectively useless. In contrast, one of the newer models, the YOLOv11, not only recovered but improved steadily, producing high-quality outlines that met clinical standards. This model was the only one that showed a consistent, smooth improvement as the data grew from forty to eight hundred images. The other newer model also improved but showed erratic behavior, getting worse before it got better as the data increased. By the time the training set reached eight hundred images, all three models had learned to perform almost identically, producing highly accurate outlines with a level of precision that would be difficult for a human to match.

The study also measured how fast these systems could process an image. The older model was slightly faster, taking about fifty-nine milliseconds to analyze a single slice, while the newest model took about seventy-two milliseconds. While the speed difference is measurable, the researchers found that the stability of the learning process was far more important than the few milliseconds saved. The most significant finding was that with the right architecture and a specific training method, a computer can learn to segment a complex bone structure with as few as one hundred annotated images. This is a crucial discovery for the medical field, where large datasets are rare, because it proves that high-quality automation is possible without needing thousands of labeled scans.

However, the researchers are careful to note the limits of their work. The entire experiment was conducted using images from a single person and a single type of MRI scanner. While the computer learned to outline the bone perfectly on the images it was trained on, it has not yet been tested on different patients, different scanners, or people with bone diseases. The high accuracy achieved in this study reflects the system's ability to learn the specific patterns of the synthetic data it was given, rather than a proven ability to generalize to the entire human population. The team concludes that while the technology is promising, the next step must be to test these systems on diverse groups of patients to ensure they work safely in real-world clinical settings. Until then, the best approach for hospitals with limited data is to choose the specific model that proved most stable in this study, rather than assuming that the newest technology will always be the most effective.

Technical Summary: Quantifying Data Efficiency in YOLO-Based Femur Segmentation for Axial T1-Weighted MRI

Problem Statement
Automated femur segmentation in T1-weighted MRI is critical for pre-operative planning, implant sizing, and biomechanical modeling. However, the clinical reality is a scarcity of large, labeled datasets, which conventional deep learning models typically require. While deep learning has improved accuracy over statistical shape models, most existing literature either focuses on large-dataset performance or investigates data efficiency without comparing architecturally distinct model families. There is a specific gap in understanding how segmentation performance scales with training data volume across different YOLO (You Only Look Once) variants, particularly in the data-scarce regime (N < 100).

Methodology
The study conducted a systematic, controlled investigation using three instance segmentation architectures from the YOLO family: YOLOv8m-seg, YOLOv11m-seg, and YOLOv26m-seg. All models utilized the medium ('m') variant to ensure comparable parameter counts and were initialized with COCO-pretrained checkpoints.

Dataset: The base data consisted of 40 T1-weighted axial MRI slices of the femur from a single healthy adult female (Siemens 3T scanner). To simulate varying data sizes, these 40 images were synthetically expanded using an anatomically validated augmentation pipeline (rotation, flipping, brightness/contrast, noise, and blur) via the Albumentations library.
Experimental Design: Five training set sizes (N = 40, 100, 200, 400, 800) were created. A fixed, unknown synthetic test set of 100 images was held out and used for evaluation across all conditions to ensure consistency.
Training Protocol: A two-phase transfer learning approach was employed to prevent catastrophic forgetting:
1. Phase 1 (20 epochs): Backbone freezing (first 10 layers) with a higher learning rate ( $10^{-3}$ ) to adapt the segmentation head.
2. Phase 2 (80 epochs): Full fine-tuning with a lower learning rate ( $10^{-4}$ ), cosine annealing, and early stopping.
Evaluation Metrics: Performance was assessed using segmentation overlap (Dice Similarity Coefficient [DSC], Intersection over Union [IoU]), boundary distance (95th-percentile Hausdorff Distance [HD95], Average Symmetric Surface Distance [ASSD]), detection metrics (mAP@0.5, F1 score), and inference speed (ms/image). Statistical significance was determined using Mann–Whitney U tests, Wilcoxon signed-rank tests, and Cohen's d effect sizes.

Key Results
The study revealed distinct performance trajectories based on dataset size and architecture choice:

Data-Scarce Regime (N = 40 to 100):
- YOLOv11m-seg demonstrated the most robust data efficiency, exhibiting a monotonically increasing DSC across all conditions. It achieved acceptable performance (DSC > 0.87) at N = 100.
- YOLOv8m-seg showed an anomalous "collapse" at N = 100, where performance dropped drastically from DSC = 0.757 (at N=40) to 0.193, before recovering at N = 200.
- YOLOv26m-seg failed at N = 40 (DSC = 0.163) but recovered rapidly at N = 100 (DSC = 0.877). However, it displayed non-monotonic behavior at N = 400.
- At N = 40, the older YOLOv8 architecture was actually the most reliable (DSC = 0.757), outperforming the newer models which struggled with overfitting or underfitting on the minimal data.
High-Data Regime (N = 800):
- All three architectures converged to statistically indistinguishable, high-performance levels (DSC $\approx$ 0.960–0.963, HD95 < 4 pixels).
- Pairwise statistical tests (Mann–Whitney) showed no significant differences between models at N = 800, indicating that architecture selection becomes negligible once sufficient data is available.
Inference Speed:
- YOLOv8m-seg remained the fastest (~~59–66 ms/image), followed by YOLOv11m-seg (~~65–67 ms/image) and YOLOv26m-seg (~70–72 ms/image).

Key Contributions and Significance
The paper's primary contribution is the quantification of the trade-off between architectural advancement and data efficiency in medical image segmentation. The authors claim that:

Architecture Selection is Critical in Low-Data Scenarios: In the data-scarce regime (N < 100), the choice of architecture is the dominant variable. Selecting the wrong model (e.g., YOLOv8 at N=100) can lead to effective model failure, whereas YOLOv11m-seg offers a reliable, monotonically improving trajectory.
Feasibility of Small Datasets: With appropriate transfer learning and anatomically relevant augmentation, YOLO-family models can achieve high segmentation accuracy (DSC > 0.85) with as few as 100 annotated slices, challenging the notion that large datasets are strictly necessary for clinical-grade performance.
Convergence at Scale: The study demonstrates that architectural differences become practically negligible at N = 800, where all models converge to similar accuracy, suggesting that for large-scale datasets, the focus may shift back to inference speed (favoring YOLOv8) rather than architectural complexity.

Limitations and Scope
The authors explicitly state that the findings are based on a single-subject, single-scanner dataset with synthetic augmentation. Consequently, the reported DSC of ~0.96 reflects performance on augmented copies of one individual's anatomy and does not yet constitute evidence of cross-population or cross-scanner generalizability. The study serves as a methods investigation into data efficiency rather than a clinical validation of a deployable tool for diverse patient populations. The mechanism behind the YOLOv8 collapse at N=100 remains a hypothesis (distributional shift, training instability, or early stopping) rather than a resolved ablation study.

Conclusion
The study concludes that YOLOv11m-seg is the recommended architecture for datasets ranging from 100 to 800 images due to its stability and monotonic improvement. YOLOv8m-seg is viable for very small datasets (N ≤ 40) but should be avoided near N = 100 due to instability. YOLOv26m-seg is viable from N = 100 onward but exhibits unpredictable intermediate behavior. The research underscores that in medical imaging contexts where data is limited, architectural selection is as important as the data itself.

Technical Summary: Quantifying Data Efficiency in YOLO-Based Femur Segmentation for Axial T1-Weighted MRI

More like this