Source-linked AI summary
nnDetection: A Self-configuring Method for Medical Object Detection
Michael Baumgartner, Paul F. Jaeger, Fabian Isensee, Klaus H. Maier-Hein
TL;DR
Medical object detection is clinically relevant, but configuring detection methods is cumbersome and iterative. nnDetection automates that configuration and achieves results superior to or on par with the state of the art across two public leaderboards and 11 proposed benchmarks.
Problem
Configuring medical object-detection methods is expert-dependent, compute-intensive, and must be repeated across tasks because data-set properties vary.
Source: §1 · p. 2Method
Following nnU-Net, nnDetection uses fixed, rule-based, and empirical choices to configure Retina U-Net automatically for medical detection problems.
Source: §1 · p. 2Results
Takeaways & Limitations
The released method, pretrained models, and benchmark annotations provide an out-of-the-box tool, a framework for method development, and a standardized comparison baseline.
Source: §4 · pp. 8–9
Abstract
from arXiv · showhide
Simultaneous localisation and categorization of objects in medical images, also referred to as medical object detection, is of high clinical relevance because diagnostic decisions often depend on rating of objects rather than e.g. pixels. For this task, the cumbersome and iterative process of method configuration constitutes a major research bottleneck. Recently, nnU-Net has tackled this challenge for the task of image segmentation with great success. Following nnU-Net's agenda, in this work we systematize and automate the configuration process for medical object detection. The resulting self-configuring method, nnDetection, adapts itself without any manual intervention to arbitrary medical detection problems while achieving results en par with or superior to the state-of-the-art. We demonstrate the effectiveness of nnDetection on two public benchmarks, ADAM and LUNA16, and propose 11 further medical object detection tasks on public data sets for comprehensive method evaluation. Code is at https://github.com/MIC-DKFZ/nnDetection .
1 Introduction
Medical object detection directly localizes and rates objects, a clinically relevant task, but configuring detection methods is complex and iterative. nnDetection adapts nnU-Net’s automated configuration approach to medical object detection and is evaluated against public benchmarks and a standardized baseline.
1 Introduction
- Detection localizes and rates entire objects, whereas segmentation categorizes pixels and may need ad-hoc postprocessing to produce object scores.That postprocessing can negatively affect performance compared with object detection methods that perform these steps during learning.Source: §1 · p. 1
- Object-detection architectures add sampling, anchor, and box-clustering choices, making configuration expert-dependent, compute-intensive, and repetitive across data sets.The repeated effort arises because medical data sets vary in their properties.Source: §1 · p. 2
- nnDetection adapts nnU-Net’s fixed, rule-based, and empirical configuration strategy to Retina U-Net using design choices developed on 10 data sets.It is presented for automatic deployment on arbitrary medical detection problems without compute beyond standard network training.Source: §1 · p. 2
- The paper evaluates nnDetection on LUNA16 and ADAM and proposes a 13-data-set benchmark comparing it with nnU-Net plus object-scoring postprocessing.The benchmark comprises tasks identified in data sets from existing segmentation challenges.Source: §1 · p. 2
- The authors release nnDetection, pretrained models, and annotations for new benchmarks as an out-of-the-box method, development framework, and standardized baseline.The stated aim is to support research interest in medical object detection.Source: §1 · p. 2
2 Methods
nnDetection configures detection pipelines through fixed, rule-based, and empirical choices. It derives rules from data-set properties, trains configured models, and tunes selected test-time parameters empirically.
2 Methods
- The design combines fixed, rule-based, and empirical choices, with a data fingerprint guiding rules and empirical parameters optimized after training.These categories and their interaction are summarized in Fig. 1.Source: §2 · p. 2, Figure 1 · p. 3
- The fixed configuration uses Retina U-Net and adaptive training sample selection, removes the anchor-center inclusion requirement, and uses 27 anchors per position.The authors report that removing the inclusion requirement avoids losing all positive anchors for small objects, while 27 anchors particularly improves results on data sets with few objects.Source: §2 · p. 4
- Rules use the data fingerprint, including object sizes, to set network topology and anchor configurations, with batch size fixed at 4.Anchor sizes are iteratively fitted on the training split; a low-resolution model is triggered for very large objects or high-resolution images.Source: §2 · p. 4
- At test time, NMS clusters predictions across overlapping inference patches, while Weighted Box Clustering combines predictions across models or test-time augmentations.Interdependent empirical parameters are initialized and optimized sequentially on the validation set; if triggered, the low-resolution model is selected using validation results.Source: §2 · pp. 4–5
- For a new data set, nnDetection extracts its fingerprint, applies rules, trains full- and, if triggered, low-resolution models with five-fold cross-validation, then ensembles five models.Configuration requires no additional compute beyond standard network training apart from the empirical choices; the authors also evaluate generalization on three additional data sets.Source: §2 · p. 5
- The nnU-Net Basic baseline converts pixel predictions to objects using connected components and maximum softmax scores; nnU-Net Plus tunes object-scoring postprocessing on training data.The Plus baseline selects thresholds, a minimum object size, and an aggregation method; the authors also report fixing numerical constants in Dice loss after observing convergence issues on LIDC.Source: §2 · p. 5
3 Experiments and Results
The experiments evaluate nnDetection on public leaderboards and a proposed multi-data-set benchmark, with comparisons against nnU-Net. Results favor nnDetection across most reported comparisons, while the authors note substantial metric volatility across data sets.
3 Experiments and Results
- The proposed benchmark uses 13 existing data sets because limited data and label noise can make evaluation metrics volatile and task-specific claims overfit.The authors argue that these concerns apply to medical object detection as well as segmentation.Source: §3 · p. 5
- 0.930 CPM puts nnDetection above previous methods on LUNA16 nodule-candidate detection; its ADAM leaderboard submission ranks third with sensitivity 0.64 at 0.3 false positives.On general LUNA16, Liu et al. [13] is the only approach reported to outperform nnDetection, using additional False Positive Reduction models.Source: Figure 2 · p. 6, §3 · p. 8
- Five segmentation-challenge data sets supply additional detection tasks; connected components produce object labels, objects under 3 mm are discarded, and results use mAP at IoU 0.1.The authors also manually remove annotations from obvious segmentation errors.Source: §3 · p. 6
- Ten data sets support development and validation, while LUNA16 and TCIA Lymph-Node form the three-data-set test pool held out from method development.The validation pool uses held-out test splits, and ADAM uses its public leaderboard as the hold-out test split.Source: §3 · p. 6
- nnDetection outperforms nnU-Net Plus on 9 of 12 cross-validation data sets and 5 of 7 hold-out splits, with substantial variation on liver and colon.Margins exceed 5% on 7 cross-validation data sets; nnU-Net Basic performs worse on 11 of 12, and four test data sets were in nnU-Net’s development pool.Source: Figure 3 · p. 7, §3 · p. 8
- Because 11 data sets lack a standardized object-detection protocol, the authors compare nnDetection with modified nnU-Net as a standardized baseline.They cite the prevalence of segmentation methods adapted to detection tasks and the limited prior study of dedicated detection methods versus segmentation substitutes.Source: §3 · pp. 6–7
4 Discussion
The authors present nnDetection as a data-agnostic approach that adapts out of the box and removes manual, iterative configuration. It performs on par with or above state of the art across public leaderboards and proposed benchmarks, and is released with resources for reuse.
4 Discussion
- nnDetection performs superior to or on par with the state of the art on two public leaderboards and 11 newly proposed benchmarks.The authors describe it as a starting point for further manual, task-specific optimization.Source: §4 · p. 8
- An additional false-positive-reduction component can further improve nnDetection’s LUNA16 results.The authors also suggest data-driven optimization of specific pipeline components as a possible route to further improvement.Source: §4 · p. 8
- The release includes models and object annotations for the new benchmarks, positioning nnDetection as an out-of-the-box tool, method-development framework, and standardized baseline.It is also presented as a benchmark for large-scale method evaluation.Source: §4 · pp. 8–9
A Supplementary Material
The supplementary material reports LUNA16 comparisons and ADAM leaderboard standing, alongside dataset annotation and split procedures. It also documents exceptions in annotation cleanup.
A Supplementary Material
- In Kits19, small clusters, often visible in only one slice, were removed from annotations; kidney annotations were excluded from training.The small-cluster removal was not applied to Decathlon Liver because it contained too many small objects.Source: Figure 4 · p. 12
- On the ADAM live leaderboard, nnDetection ranks third with sensitivity 0.64 at a false-positive count of 0.3.The leaderboard’s top-left corner is best.Source: Figure 5 · p. 12
- The ADAM examples show ground-truth aneurysm annotations beside nnDetection predictions; small aneurysms have little tolerance for predictions to count as true positives.Their small radius makes the live leaderboard’s true-positive criterion particularly strict.Source: Figure 5 · p. 12
- For datasets without object labels, connected components were derived from segmentation annotations, objects under 3 mm were discarded, and obvious errors were manually checked.For datasets with additional organ annotations, only tumor labels were used for training; ADAM used a patient-stratified split and LIDC a custom split.Source: Table 1 · p. 13
- nnDetection outperforms methods without an additional False Positive Reduction stage; only Liu et al. [13] with FPR achieves a higher CPM score.Liu et al.’s higher CPM is especially associated with improved performance at low false-positive-per-scan thresholds.Source: Table 2 · p. 13