Title: Characterizing 3D Kidney Lesions via Multi-Granularity Analysis of CT Volumes

Abstract:

While radiology reports typically detail kidney lesions based on their type, dimensions, enhancement patterns, and attenuation levels, current 3D analytical methods are limited to making predictions at the organ or patient level. To address this gap, we reframe kidney CT characterization as a per-lesion set-prediction problem, employing a single model capable of outputting a variable quantity of lesions per kidney, each accompanied by four distinct clinical attributes. Our study utilizes a curated dataset of 2,619 CT volumes from 788 patients at a single academic medical center, featuring both side-level and per-lesion labels for multi-granularity analysis. For zero-shot external validation, we incorporated the KiTS23 dataset, comprising 489 cases.

We introduce LesionDETR, a DETR-style architecture that incorporates size-distance Hungarian matching and a hierarchical loss function designed to aggregate per-slot outputs toward side-level objectives. Our evaluation across four input representations and six encoder initializations highlights two critical design factors: the inclusion of a segmentation mask as an input channel and the use of same-domain abdominal pretraining (SuPreM). Notably, generic pretraining on large corpora offered no advantage over random initialization. On the UF-Health dataset, LesionDETR achieved a bilateral side-level abnormality AUC of $0.799 \pm 0.009$, while reaching $0.817 \pm 0.072$ on KiTS23.

A count-conditioned variant of the model attained a per-lesion mean Average Precision (mAP) of $0.190 \pm 0.083$ for cystic lesions. However, the Average Precision (AP) for rare solid lesions remained at the noise floor, suggesting that future improvements will depend more on targeted data collection than on architectural changes. Ultimately, this framework provides verified per-lesion predictions suitable for downstream structured report generation.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC