2477 - LymphLoc: Localization of Mediastinal Lymph Nodes Using an Anatomic Atlas Reference for Downstream Contouring

Purpose/Objective(s):

Locating target structures using AI can serve as an intermediate step towards automated radiotherapy contouring. Specific guidelines for target structure localization can be extracted from training materials, such as an anatomical atlas. We investigate the use of off-the-shelf vision-language models paired with a training atlas and representative examples for the localization of mediastinal lymph nodes.

Materials/Methods:

We use the LyNoS mediastinal lymph node segmentation dataset, for which we group the stations as follows: 1L, 1R, 2L, 2R, 3P, 3A, 4L, 4R, 5, 6, 7, 8, 9, 10-12L and 10-12R. For few-shot examples for each mediastinal nodal station, 31 cases were used. For testing, 73 cases, containing 639 stations were used. We prompt the VLM, GPT 5.2, to locate a specific station in a provided axial CT slice. As a visual reference, we provide normalized coordinate axes (x and y coordinates 0-1). As prior anatomical knowledge, the mediastinal lymph node atlas paper by Sarrut et al. (2013) is provided to the model. This is a paper describing the extent of mediastinal lymph node stations, paired with illustrative figures. Additionally, we give the VLM 1, 2, 4 or 6 few-shot examples before querying the model with a new image. We evaluate the Hausdorff distance between the predicted point and a ground truth mask.

Results:

Introducing an atlas significantly improves average Hausdorff distance. Averaged over all few-shot settings (Table 1), an average Hausdorff distance of 32.80 mm is achieved with an atlas and 37.31 mm without (Wilcoxon signed rank p < 0.05). Subgroup analysis shows this difference is also significant for 0, 1, 2 and 4 few shot example subgroups, but not for the 6 few – shot example subgroup, tested with the Wilcoxon signed rank test. Adding a normalized numbered axes overlay improves Hausdorff distance performance from 37.87 ±19.09 to 33.50 ± 18.67 mm on a test with atlas but without few-shot examples. Combining an atlas, few-shot examples and a normalized numbered axes overlay yields the best overall performance.

Conclusion:

Generalist vision language models struggle with radiological localization tasks. Leveraging training tools used by radiation oncology trainees, such as a mediastinal lymph node atlas and examples, informs the localization prediction task and improves localization performance. However, performance remains unsatisfactory for use in downstream segmentation steering pipelines. Future work should focus on fine-tuning vision language models for medical localization.

# few-shot	Atlas	No Atlas	P
0	33.50 ±18.67	39.24 ±21.73	< 0.001
1	32.89 ±18.75	36.67 ±20.92	< 0.001
2	32.64 ±18.21	36.99 ±21.06	< 0.001
4	32.36 ±18.86	37.40 ±21.76	< 0.001
6	32.60 ±19.05	36.25 ±21.74	0.063

Table 1: The effect of adding atlas information across different amounts of few shot examples. Tested with normalized coordinate axes overlays.