Computer-Aided Tumor Diagnosis of Microscopy Images

Aubreville M (2020)

Publication Language: English

Publication Type: Thesis

Publication year: 2020



For the diagnosis of medical images, computer-aided methods can help to lower time requirements and reduce human error by highlighting regions of particular interest, by aiding inexperienced physicians in their interpretation, or just by providing new visualizations using secondary data. In this work, several new methods to support medical experts in the diagnosis of microscopy images are presented. The first part of the thesis focuses on the processing of images acquired using confocal laser endomicroscopy. This relatively novel imaging modality provides in-vivo, in-situ microscopy at very high magnification and hence allows for microstructural analysis of tissue. A common problem with this imaging modality is the occurrence of image deteriorations. Motion artifacts are amongst the most common of these and are especially challenging to detect algorithmically. This thesis investigates multiple methods for the detection of these impediments within the image. Besides a classical machine learning pipeline, involving features known from literature and novel, specially designed features for motion artifact detection, a custom deep learning architecture that achieved an accuracy of 94% and outperformed traditional approaches is introduced. Confocal laser endomicroscopy can be used, amongst other relevant questions, for the diagnosis of squamous cell carcinoma (a tumor of the epithelium). This malignancy has a high prevalence in the head and neck region and often occurs in the upper respiratory and digestive tract. Automatic detection of these tumors using non-invasive methods could enable usage for screening by medical personnel inexperienced in the modality and thus provide possibilities for earlier detection, which could improve therapeutical outcomes. In this thesis, new, deep learning-based methods are presented and shown to achieve classification accuracies of greater than 90%, which is similar to human experts on the same modality. The second part of this thesis is about the detection of cells undergoing cell division (mitotic figures) in hematoxylin- and eosin-stained bright-field microscopy images. While it is a highly relevant task with many algorithmic competitions held on the topic, it is far from being solved. We were able to show that one crucial factor limiting the development of clinical-grade solutions is the availability of data sets of sufficient quantity and quality. In this thesis, we present new methods to efficiently create datasets for microscopy cell annotations and demonstrate their capabilities by introducing a newly created data set of unprecedented size that the methods allowed us to build. Using detection algorithms trained using the data set, it was, for the first time, possible to perform mitotic figure detection on completely labeled whole slide images of tumor tissue. In the technical validation of the data set, which was built using specimens from canine cutaneous mast cell tumors, an F1-score of 0.82 was found. An essential part of tumor grading is the selection of the mitotically most active area of tissue. Performing the mitotic count in this area is especially meaningful since the corresponding mitotic activity is known to be strongly correlated with tumor proliferation and is thus highly relevant for prognosis. This task was never before assessed algorithmically due to the lack of suitable data sets. Three novel approaches were described and compared for this purpose and shown to outperform human experts significantly. Finally, the work investigates domain-transfer and multi-domain applications of the generated models to a broader range of tissue types and species. When training and cross-domain-evaluating on two further novel whole slide image data sets of canine mammary carcinoma and human meningioma, a substantial domain shift was found. Besides the exploitation of this shift for detection, the work describes and evaluates new algorithmic approaches for unsupervised domain adaptation.

Authors with CRIS profile

How to cite


Aubreville, M. (2020). Computer-Aided Tumor Diagnosis of Microscopy Images (Dissertation).


Aubreville, Marc. Computer-Aided Tumor Diagnosis of Microscopy Images. Dissertation, 2020.

BibTeX: Download