Directly to content
  1. Publishing |
  2. Search |
  3. Browse |
  4. Recent items rss |
  5. Open Access |
  6. Jur. Issues |
  7. DeutschClear Cookie - decide language by browser settings

From Manual to Automated Design of Biomedical Semantic Segmentation Methods

Isensee, Fabian

[img]
Preview
PDF, English - main document
Download (30MB) | Terms of use

Citation of documents: Please do not cite the URL that is displayed in your browser location input, instead use the DOI, URN or the persistent URL below, as we can guarantee their long-time accessibility.

Abstract

Digital imaging plays an increasingly important role in clinical practice. With the number of images that are routinely acquired on the rise, the number of experts devoted to analyzing them is by far not increasing as rapidly. This alarming disparity calls for automated image analysis methods to ease the burden on the experts and prevent a degradation of the quality of care. Semantic segmentation plays a central role in extracting clinically relevant information from images, either all by themselves or as part of more elaborate pipelines, and constitutes one of the most active fields of research in medical image analysis. Thereby, the diversity of datasets is mirrored by an equally diverse number of segmentation methods, each being optimized for the datasets they are addressing. The resulting diversity of methods does not come without downsides: The specialized nature of these segmentation methods causes a dataset dependency which makes them unable to be transferred to other segmentation problems. Not only does this result in issues with out-of-the-box applicability, but it also adversely affects future method development: Improvements over baselines that are demonstrated on one dataset rarely transfer to another, testifying a lack of reproducibility and causing a frustrating literature landscape in which it is difficult to discern veritable and long lasting methodological advances from noise.

We study three different segmentation tasks in depth with the goal of understanding what makes a good segmentation model and which of the recently proposed methods are truly required to obtain competitive segmentation performance. To this end, we design state of the art segmentation models for brain tumor segmentation, cardiac substructure segmentation and kidney and kidney tumor segmentation. Each of our methods is evaluated in the context of international competitions, ensuring objective performance comparison with other methods. We obtained the third place in BraTS 2017, the second place in BraTS 2018, the first place in ACDC and the first place in the highly competitive KiTS challenge. Our analysis of the four segmentation methods reveals that competitive segmentation performance for all of these tasks can be achieved with a standard, but well-tuned U-Net architecture, which is surprising given the recent focus in the literature on finding better network architectures. Furthermore, we identify certain similarities between our segmentation pipelines and notice that their dissimilarities merely reflect well-structured adaptations in response to certain dataset properties. This leads to the hypothesis that we can identify a direct relation between the properties of a dataset and the design choices that lead to a good segmentation model for it.

Based on this hypothesis we develop nnU-Net, the first method that breaks the dataset dependency of traditional segmentation methods. Traditional segmentation methods must be developed by experts, going through an iterative trial-and-error process until they have identified a good segmentation pipeline for a given dataset. This process ultimately results in a fixed pipeline configuration which may be incompatible with other datasets, requiring extensive re-optimization. In contrast, nnU-Net makes use of a generalizing method template that is dynamically and automatically adapted to each dataset it is applied to. This is achieved by condensing domain knowledge about the design of segmentation methods into inductive biases. Specifically, we identify certain pipeline hyperparameters that do not need to be adapted and for which a good default value can be set for all datasets (called blueprint parameters). They are complemented with a comprehensible set of heuristic rules, which explicitly encode how the segmentation pipeline and the network architecture that is used along with it must be adapted for each dataset (inferred parameters). Finally, a limited number of design choices is determined through empirical evaluation (empirical parameters). Following the analysis of our previously designed specialized pipelines, the basic network architecture type used is the standard U-Net, coining the name of our method: nnU-Net (”No New Net”). We apply nnU-Net to 19 diverse datasets originating from segmentation competitions in the biomedical domain. Despite being applied without manual intervention, nnU-Net sets a new state of the art in 29 out of the 49 different segmentation tasks encountered in these datasets. This is remarkable considering that nnU-Net competed against specialized manually tuned algorithms on each of them. nnU-Net is the first out-of-the-box tool that makes state of the art semantic segmentation methods accessible to non-experts. As a framework, it catalyzes future method development: new design concepts can be implemented into nnU-Net and leverage its dynamic nature to be evaluated across a wide variety of datasets without the need for manual re-tuning.

In conclusion, the thesis presented here exposed critical weaknesses in the current way of segmentation method development. The dataset dependency of segmentation methods impedes scientific progress by confining researchers to a subset of datasets available in the domain, causing noisy evaluation and in turn a literature landscape in which results are difficult to reproduce and true methodological advances are difficult to discern. Additionally, non-experts were barred access to state of the art segmentation for their custom datasets because method development is a time consuming trial-and-error process that needs expertise to be done correctly. We propose to address this situation with nnU-Net, a segmentation method that automatically and dynamically adapts itself to arbitrary datasets, not only making out-of-the-box segmentation available for everyone but also enabling more robust decision making in the development of segmentation methods by enabling easy and convenient evaluation across multiple datasets.

Item Type: Dissertation
Supervisor: Brors, Prof. Dr. Benedikt
Place of Publication: Heidelberg
Date of thesis defense: 24 September 2020
Date Deposited: 05 Feb 2021 18:57
Date: 2020
Faculties / Institutes: The Faculty of Bio Sciences > Dean's Office of the Faculty of Bio Sciences
Subjects: 004 Data processing Computer science
570 Life sciences
About | FAQ | Contact | Imprint |
OA-LogoDINI certificate 2013Logo der Open-Archives-Initiative