Open Access

Dominik Müller

• In recent years, deep neural network based medical image analysis has become quite powerful and achieved similar results performance-wise as experts. Consequently, the integration of these tools into the clinical routine as clinical decision support systems is highly desired. The benefits of automatic image analysis for clinicians are massive, ranging from improved diagnostic as well as treatment quality to increased time-efficiency through automated structured reporting. However, implementations in the literature revealed a significant lack of standardization in pipeline building resulting in low reproducibility, high complexity through extensive knowledge requirements for building state-of-the-art pipelines, and difficulties for application in clinical research. The main objective of this work is the standardization of pipeline building in deep neural network based medical image segmentation and classification. This is why the Python frameworks MIScnn for medical imageIn recent years, deep neural network based medical image analysis has become quite powerful and achieved similar results performance-wise as experts. Consequently, the integration of these tools into the clinical routine as clinical decision support systems is highly desired. The benefits of automatic image analysis for clinicians are massive, ranging from improved diagnostic as well as treatment quality to increased time-efficiency through automated structured reporting. However, implementations in the literature revealed a significant lack of standardization in pipeline building resulting in low reproducibility, high complexity through extensive knowledge requirements for building state-of-the-art pipelines, and difficulties for application in clinical research. The main objective of this work is the standardization of pipeline building in deep neural network based medical image segmentation and classification. This is why the Python frameworks MIScnn for medical image segmentation and AUCMEDI for medical image classification are proposed which simplify the implementation process through intuitive building blocks eliminating the need for time-consuming and error-prone implementation of common components from scratch. The proposed frameworks include state-of-the-art methodology, follow outstanding open-source principles like extensive documentation as well as stability, offer rapid as well as simple application capabilities for deep learning experts as well as clinical researchers, and provide cutting-edge high-performance competitive with the strongest implementations in the literature. As secondary objectives, this work presents more than a dozen in-house studies as well as discusses various external studies utilizing the proposed frameworks in order to prove the capabilities of standardized medical image analysis. The presented studies demonstrate excellent predictive capabilities in applications ranging from COVID-19 detection in computed tomography scans to the integration into a clinical study workflow for Gleason grading of prostate cancer microscopy sections and advance the state-of-the-art in medical image analysis by simplifying experimentation setups for research. Furthermore, studies for increasing reproducibility in performance assessment of medical image segmentation are presented including an open-source metric library for standardized evaluation and a community guideline on proper metric usage. The proposed contributions in this work improve the knowledge representation of the field, enable rapid as well as high-performing applications, facilitate further research, and strengthen the reproducibility of future studies.…