Manual segmentation, however, is a labor-intensive and time-consuming process. While machine learning (ML) models can assist in automating segmentation, they typically require extensive training datasets to achieve accurate results. In situations where such data is unavailable, manual segmentation remains the only viable option. This challenge is effectively addressed by Interactive Machine Learning (IML).
IML utilizes a lightweight model trained on a limited dataset. Initially, the model may produce suboptimal segmentations when tasked with identifying a specific object or organ. The user, such as a radiologist, can refine these results by providing corrective inputs, such as scribbles. The IML framework adapts to these corrections, enabling the model to progressively improve its accuracy. Over time, the model achieves segmentation quality comparable to that of a foundation model.
Additionally, the structural similarity of organs across consecutive frames within a medical volume can be leveraged. Once a radiologist refines the segmentation for a single frame, Video Object Segmentation techniques can be employed to propagate the segmentation mask across the remaining frames. In cases where inaccuracies occur during this propagation, users can make further adjustments through corrective inputs, ensuring precise segmentation throughout the volume.
