Learning Beyond Labels: Understanding the Role of Imperfect Supervision in Medical Image Segmentation

Abstract

Medical image segmentation often relies either on traditional image processing methods, which can be limited in their ability to handle complex image characteristics, or on deep learning approaches, which require large amounts of high-quality annotated data. Hybrid methods that combine these two paradigms offer a potential way to reduce annotation requirements while maintaining strong segmentation performance. This thesis investigates such a hybrid approach by combining a gradient-based wavefront propagation method, namely the fast marching method, with a deep learning model.

The central objective is to analyze how imperfect annotations generated by the fast marching method propagate through the deep learning model, and to assess how much and what type of supervision is sufficient to achieve reliable segmentation performance. To enable a systematic and controlled investigation, CT-like volumes of the myenteric plexus with known two-class ground truth are simulated. The fast marching method is then applied to generate segmentation masks under different conditions, including deliberately undersegmented and oversegmented configurations, which are subsequently used as imperfect supervision to train an Attention Res-UNet with Guided Decoder.

The results show that deep learning models trained on the fast marching–derived masks can substantially reduce false negatives recovering large missing regions, even when the annotations are highly incomplete. However, false positives tend to persist and may be amplified by the deep learning model. Overall, the findings demonstrate that fast marching can serve as an effective pre-processing step in a hybrid segmentation pipeline, provided that systematic oversegmentation is avoided. The work highlights the importance of understanding the error characteristics of supervisory masks and contributes to a more nuanced view of imperfect supervision in medical image segmentation.

Popular Abstract

Accurate segmentation of medical images is an important but challenging task in healthcare and medical research. Traditionally, this has been done either by manual annotation, which is time-consuming, or by automated methods that rely on carefully tuned rules and parameters. More recently, deep learning has achieved impressive results, but at the cost of requiring large amounts of manually annotated training data, which are often difficult to obtain.

This thesis explores a hybrid approach that combines a traditional segmentation technique, the fast marching method, with a deep learning model. The idea is to use fast marching to generate rough, imperfect segmentations that can then be refined by deep learning, reducing the need for fully annotated data. To study this in a controlled way, synthetic CT-like images of the myenteric plexus were generated, allowing the true segmentation to be known exactly.

The experiments show that the deep learning model can successfully learn from imperfect fast marching segmentations and often improves them substantially by filling in missing regions. In particular, when the fast marching method misses parts of the target structure, the deep learning model is able to recover much of this information. However, when the fast marching method includes regions that do not belong to the target structure, these errors tend to remain and may even become more pronounced.

These results highlight that not all imperfect annotations are equally useful: missing information is often easier for a deep learning model to correct than incorrect information. The study suggests that hybrid segmentation approaches can be a promising way to reduce annotation effort, but that careful attention must be paid to how the initial segmentations are generated. In the future, such methods could help make medical image analysis more efficient and adaptable in clinical practice.

• author: Parkkola, Sara Lotta Alexandra
• supervisor: Ted Kronvall
• organization: Mathematical Statistics
• course: MASM02 20252
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: medical image segmentation, fast marching method, deep learning, neural network, Attention Res-UNet, imperfect supervision, hybrid segmentation, CT imaging, myenteric plexus
• publication/series: Master’s Theses in Mathematical Sciences
• report number: LUNFMS-3137-2026
• ISSN: 1404-6342
• other publication id: 2026:E16
• language: English
• id: 9223791
• date added to LUP: 2026-04-09 15:31:37
• date last changed: 2026-04-09 15:31:37