💻 computer science

Advanced Tumor Segmentation in PET/CT Imaging: A Training Strategy Study with nnU-Net for AutoPET III

This paper presents a third-place ranking AutoPET III challenge solution that leverages the nnU-Net framework with a ResNet encoder and optimized training strategies, including CraveMix augmentation and batch dice loss, to achieve robust whole-body tumor segmentation in PET/CT imaging with a Dice score of 0.80.

Original authors: Hussain Alasmawi

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hussain Alasmawi

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to teach a robot to find tumors in 3D pictures of the human body. These pictures are like complex maps made from two different types of cameras: one that sees the body's structure (CT) and one that sees where the cancer is "glowing" (PET). This is a tough job because tumors come in all shapes, sizes, and locations, and sometimes the pictures look very different from patient to patient.

This paper is a report from a competition called AutoPET III, where teams tried to build the best robot to do this job automatically. The author, Hussain Alasmawi, didn't try to invent a brand-new robot from scratch. Instead, he took a very smart, pre-made robot called nnU-Net (think of it as a high-end, versatile camera drone) and experimented with different ways to train it.

Here is how the training experiments worked, using simple analogies:

The Training "Recipes"

The author tested four different "recipes" to see which one made the robot the best at spotting tumors without making mistakes.

The Standard Recipe (Baseline): This was the default setting. It's like giving the robot the standard instruction manual. It works okay, but maybe not perfectly.
The "Z-Score" Recipe: This changed how the robot reads the brightness of the images. Imagine if you were trying to find a faint star in the sky; you might adjust your telescope to ignore the bright moon and focus on the dimmer stars. This recipe helped the robot ignore "noise" and stop seeing things that weren't tumors (reducing False Positives).
The "BraTS 2020" Recipe: This was inspired by a previous winner in a brain tumor contest. Instead of teaching the robot to look at one patient at a time, this method made it look at a whole group of patients at once (a larger "batch"). It's like a teacher grading a whole class's test together to see the general pattern of mistakes, rather than grading one student at a time. This helped the robot handle the fact that some tumors are tiny and others are huge.
The "CraveMix" Recipe: This was a data augmentation trick. Imagine you are teaching a child to recognize dogs. You show them pictures of dogs, but then you also show them pictures where the dog is wearing a hat, or is upside down, or is in the rain. This recipe created fake, slightly altered images to make the training data more diverse, hoping the robot would become more flexible.

The Results: Finding the Sweet Spot

The author ran these experiments and found that there is no single "perfect" recipe. Each one had a different strength and weakness, like choosing between a sports car and a truck:

The Standard and Z-Score recipes were good at finding the tumors overall (high accuracy score), but the Z-Score recipe was particularly good at not crying "Wolf!" when there was no wolf (very few false alarms). However, it sometimes missed the smaller wolves (more missed tumors).
The BraTS 2020 recipe was the best at being robust. It didn't get confused by the different sizes of tumors and made very few false alarms. It was the most reliable when facing new, unseen patients.
The CraveMix recipe helped a little bit, but it didn't make a huge difference in stopping the robot from making mistakes.

The Final Outcome

Because the competition rules only allowed the author to submit two versions of the robot, he chose the BraTS 2020 and Z-Score recipes. He combined their strengths to create the final entry.

The Verdict:
By simply tweaking how the robot was trained—without changing the robot's actual brain or structure—the author's team achieved a very high score. Their method was good enough to win 3rd place in the entire AutoPET III challenge.

The Big Takeaway:
The paper proves that you don't always need to build a fancier, more complex machine to get better results. Sometimes, just changing the training strategy (like adjusting the lighting, the group size, or the practice exercises) is enough to make a huge difference in how well the AI performs.

Technical Summary: Advanced Tumor Segmentation in PET/CT Imaging

Problem Statement
Whole-body tumor segmentation in PET/CT imaging is essential for precise disease evaluation and treatment planning but remains a significant challenge due to the high variability in lesion size, contrast, and anatomical distribution. Manual segmentation is not only time-consuming but also prone to intra- and inter-observer variability. The specific context of this work is the AutoPET III challenge, which focuses on developing models capable of generalizing across different radiotracers (FDG and PSMA) and multi-center data, a requirement that standard segmentation approaches often struggle to meet.

Methodology
The authors utilize the nnU-Net framework as a baseline, specifically employing a Residual Encoder UNet variant with a ResNet-based encoder to enhance feature representation and gradient flow. The study does not propose novel architectural changes; instead, it systematically investigates the impact of various training strategies on a fixed architecture (Residual Encoder Large configuration with a $192 \times 192 \times 192$ voxel patch size).

Key methodological components include:

Dataset: The study uses the AutoPET-III dataset, comprising whole-body PET/CT scans from two medical centers. The data includes two primary cohorts: FDG (501 cancer patients, 513 negative controls) and PSMA (537 cases with PSMA-avid lesions, 60 without).
Inference Strategy: To handle large whole-body volumes, a sliding window inference strategy is employed. Robustness is enhanced through model ensembling, where final predictions are derived by averaging probabilities from models trained via 5-fold cross-validation.
Training Strategy Experiments: Four distinct configurations were evaluated:
1. Baseline: Default nnU-Net configuration with standard CT normalization and a $192^3$ patch size.
2. Z-score Normalization: Replaced default PET normalization with Z-score normalization.
3. BraTS 2020 Strategy: Inspired by the BraTS 2020 winning approach, this involved increasing the batch size from 2 to 5 and utilizing batch Dice loss (computing Dice scores across the entire batch rather than per sample) to better handle tumor size variability.
4. CraveMix Augmentation: Integration of CraveMix data augmentation to generate 350 synthetic samples per fold.

Key Results
The experiments, evaluated via 5-fold cross-validation and on a preliminary test phase, revealed distinct trade-offs between different strategies:

Baseline & Z-score: Both achieved the highest cross-validation Dice scores (0.68). However, Z-score normalization significantly reduced false positives (FP) in the test phase to 4.09 (compared to 16.33 for the baseline), though this came at the cost of a higher false negative (FN) rate (18.74).
BraTS 2020 Strategy: This approach yielded a lower cross-validation Dice (0.66) but demonstrated superior generalization on the test phase with a Dice score of 0.79. Notably, it achieved the lowest false positive rate of all strategies (1.19), indicating improved robustness to lesion variability.
CraveMix: This strategy provided moderate improvements (Test Dice 0.78) but did not significantly reduce false positives or false negatives compared to other methods.

Based on these results, the final submission for the AutoPET III challenge utilized the BraTS 2020 and Z-score normalization models. The combined method achieved a Dice score of up to 0.80 on the preliminary test phase and secured third place in the challenge.

Significance and Claims
The paper claims that significant performance gains in medical image segmentation can be achieved without modifying the underlying model architecture. Instead, the study emphasizes that training protocol design—specifically the choice of normalization schemes, batch-level optimization (batch Dice loss), and data augmentation—is critical.

The authors highlight that different strategies lead to distinct performance trade-offs:

Z-score normalization is effective for reducing false positives.
Batch Dice optimization (BraTS 2020 strategy) enhances generalization and robustness to lesion variability.

The work demonstrates that systematic investigation of training strategies within the nnU-Net framework is a viable path to competitive performance in complex, multi-center, multi-tracer segmentation tasks. The code for the proposed method is made publicly available.

The Training "Recipes"

The Results: Finding the Sweet Spot

The Final Outcome

Technical Summary: Advanced Tumor Segmentation in PET/CT Imaging

More like this