Forschungsthemen
[BA] Software Design for Semi-Supervised Learning for Bone Marrow Cell Classification
Cytology is a diagnostic tool used for cancer and other diseases. The cells in mi- croscope images are often clustered, making it difficult to accurately identify them. Develop- ing a pipeline based on deep learning and convolutional neural networks can help computers identify cell types more accurately and efficiently. However, building and training such a CNN is a challenging task, especially as pathology image data used for training is largely unlabeled. Self-supervised learning and self-training can provide solutions to this problem. By inputting a whole-slide image or an already segmented cell image, a trained model can rapidly and accu- rately identify the type of cell and provide a basis for the doctor to make a correct diagnosis. The ability to identify cell types accurately is an essential skill for doctors, as it plays a critical role in cancer diagnosis. Despite the large volume of pathology image data available, it takes a long time to train a competent cytologist, and different doctors may come to different conclusions when identifying cell types. Moreover, the process of confirming cells in front of the micro- scope is time-consuming and energy-intensive. Therefore, the development of deep learning models can be a promising approach to improve cytology accuracy and efficiency. This thesis aims to investigate the use of a CNN model based on self-supervised learning to achieve faster and better cell classifications for cancer diagnosis and other screening and diagnostic areas.
Betreuer: Karsten Wendt