PaperlayerView exact revision on arXiv

Source-linked AI summary

Weakly-supervised Kidney Tumor Classification from CT Scans with Multi-Instance Learning and Anatomical Filtering

Joonas Ariva, Dmytro Fishman

arXiv:2609.07178v1cs.CV

TL;DR

Precise pixel-level CT annotations are costly, while scan-level labels reduce annotation demands but leave low supervision ratios and large-volume training challenges. This paper combines Compass anatomical filtering with multi-instance learning for weakly supervised kidney tumor classification. The best models reach F1 = 0.83 on the held-out TUH test set, and Compass filtering is important for ABMIL's out-of-distribution generalization.

  • Problem

    Pixel-level CT labels require substantial radiologist effort, while scan-level classification labels provide a lower supervision ratio and large CT volumes pose training challenges.

    Source: §1 · p. 2

  • Method

    The method combines self-supervised Compass anatomical filtering with MIL over 2D CT slices, using scan-level labels for classification and no pixel-level segmentation masks.

    Source: §1 · pp. 2–3, §2.2 · p. 5

  • Results

    F1 = 0.83 on the held-out TUH test set, and Compass filtering is critical to preventing poor out-of-distribution generalization by ABMIL.

    Source: §5 · p. 12, §6 · p. 13

  • Takeaways & Limitations

    The results support weakly supervised MIL for kidney tumor classification as a proof of concept for a broader pipeline that could be applied to other organs and pathologies.

    Source: §1 · p. 3, §6 · p. 13

    Because KiTS and KiRC contain no control scans, external evaluation is limited to sensitivity and cannot establish specificity or precision.

    Source: §5 · p. 13

Abstract

from arXiv · show

Deep learning models for CT scan analysis are often limited by the scarcity of precise pixel-level annotations, which require significant radiologist effort to produce. Training on scan-level labels alone reduces annotation requirements but introduces challenges: low supervision ratios and large input volumes make models prone to overfitting and shortcut learning. In this work, we investigate two complementary methods to address these challenges: multi-instance learning (MIL) and anatomical filtering. MIL divides CT volumes into 2D slice instances, enabling efficient 2D architectures with ImageNet pretraining rather than computationally demanding 3D models. Anatomical filtering uses Compass, our self-supervised body part regression model, to crop scans to pathology-relevant subregions without requiring segmentation masks. We evaluate two MIL frameworks - Attention-based MIL (ABMIL) and FocusMIL - on kidney tumor classification across one internal dataset (TUH) and two external datasets (KiTS23 and TCGA-KiRC). Our best models achieve F1 = 0.83 on the internal test set using only scan-level labels. We further show that anatomical filtering with the Compass model is critical for the out-of-distribution generalization of embedding-based ABMIL, while instance-based FocusMIL demonstrates greater inherent robustness to distribution shift. While evaluated on kidney tumors, we consider this a proof-of-concept for a broader weakly supervised CT classification pipeline applicable to other organs and pathologies.

1 Introduction

The paper addresses the annotation burden and training challenges of CT classification by combining multi-instance learning with Compass-based anatomical filtering. It evaluates this weakly supervised approach for kidney tumor classification and frames it as a proof of concept for other organs and pathologies.

1 Introduction

  • Pixel-level labels are costly because they require radiologist effort, while scan-level labels reduce annotation needs but provide a lower supervision ratio and can leave models prone to overfitting.Large CT volumes also create memory demands that can exceed available GPU capacity.Source: §1 · p. 2
  • Compass uses self-supervised body part regression to filter scans to pathology-relevant regions without segmentation masks, while multi-instance learning divides volumes into 2D slice instances.Together, these methods target training efficiency and classification performance.Source: §1 · pp. 2–3
  • MIL enables efficient 2D architectures with ImageNet pretraining instead of computationally demanding 3D models.Reduced input dimensionality may also reduce overfitting to spurious scan-level features.Source: §1 · p. 3
  • The study evaluates two MIL frameworks against a whole-scan baseline and combines them with Compass; it presents kidney tumor classification as a proof of concept for broader applications.The authors state that, to their knowledge, prior work had not applied MIL to kidney tumor classification.Source: §1 · p. 3

2 Background

The background introduces multi-instance learning and body part regression as ways to work with scan-level labels and locate relevant anatomy. It also situates the work against prior CT applications that commonly rely on segmentation masks.

2.1 Multi-instance learning

  • Multi-instance learning divides an image into smaller instances grouped as a bag, which is positive when one or more instances are positive under the standard MIL assumption.This reduces input dimensionality and permits bag-level supervision without instance-level labels.Source: §2.1 · p. 3
  • Instance-based MIL pools instance-level scores, whereas embedding-based MIL pools feature vectors before bag-level classification; this study uses FocusMIL with max pooling and ABMIL with attention pooling.These approaches differ in whether scoring or pooling comes first.Source: §2.1 · pp. 3–4, Figure 1 · p. 4

2.2 MIL and weak supervision in CT imaging

  • Earlier CT MIL applications covered thoracic and abdominal pathologies, but the authors state that MIL had not previously been applied to kidney tumor classification.Prior approaches commonly relied on segmentation masks for region-of-interest preprocessing.Source: §2.2 · p. 5
  • Prior CT MIL work commonly used organ or lesion segmentation masks to define regions of interest, whereas this study uses scan-level training labels and sparse slice-boundary annotations without pixel-level masks.The authors describe this as a weaker supervision regime that could support application to other organs or pathologies.Source: §2.2 · p. 5

2.3 Body part regression

  • Body part regression maps CT slices to continuous anatomical coordinates using slice order as self-supervision; the authors develop Compass and examine its effect on kidney tumor classification.They state that prior work had not examined BPR's effect on downstream pathology classification.Source: §2.3 · p. 5

3 Methods

The methods train Compass for anatomical filtering and then train MIL classifiers on the reduced scans. The experiments compare two MIL designs with a 3D baseline across internal and external kidney tumor datasets.

3 Methods

  • Training proceeds in two phases: Compass is trained, calibrated for the kidney region, and used to filter the dataset before a MIL classifier is trained on the reduced scans.Calibration uses two slice-level boundary annotations per scan to select the slices retained for classification.Source: §3 · p. 6, §3.1 · p. 7

3.1 Compass model

  • Compass maps CT slices to scalar scores and is trained by matching predicted pairwise slice distances to true distances, encouraging scores proportional to anatomical distance.Calibration thresholds are set from kidney-containing boundary slices in a small annotated scan set.Source: §3.1 · pp. 6–7

3.2 Classification models

  • The comparison includes a 3D ResNet18 baseline, attention-based ABMIL, and max-pooling FocusMIL; both MIL models use ImageNet-pretrained ResNet18 backbones.FocusMIL adds a variational information bottleneck with KL loss to regularize instance representations.Source: §3.2 · p. 7

3.3 Datasets

  • The TUH dataset contains 363 malignant-tumor scans and 723 controls, with a held-out test set of 100 cases and 100 controls; KiTS23 and TCGA-KiRC provide external tumor-positive evaluations limited to recall.Pixel-level labels were reserved for evaluation and Compass calibration, while classification used scan-level labels.Source: §3.3 · pp. 7–8

3.4 Implementation details

  • Scans were resampled to 2 mm isotropic spacing, cropped and padded to 250 × 250 pixels, soft-tissue windowed, and normalized; training used specified distributed GPU and optimization settings.For 2D MIL, each slice was stacked with its two immediate axial neighbours as a three-channel input.Source: §3.3 · p. 8, §3.4 · p. 8

4 Experiments & Results

Compass filtering retained tumor-containing slices while reducing scan size, and the classification experiments found strong internal performance with different external-generalization behavior for ABMIL and FocusMIL.

4.1 Effects of Compass model filtering

  • Tumorous-slice recall on TUH rose from 0.87 with two calibration scans to 0.97 with 200, reaching 0.95 with ten scans.Filtering reduced TUH scan size by 58%–68%; no experiment removed all tumor-containing slices from any scan.Source: §4.1 · p. 9, Figure 3 · p. 10

4.2 Classification experiments

  • The experiments evaluated models trained on TUH against its test set and the external KiTS and KiRC datasets, with results averaged across three runs.The reported epoch was selected by the highest TUH test F1; frozen-backbone experiments performed poorly, so reported models used end-to-end fine-tuning.Source: §4.2 · pp. 9–10
  • Without filtering, ABMIL recall was 0.36 on KiTS and 0.55 on KiRC; filtering raised these to 0.59 and 0.78, while FocusMIL was less affected.FocusMIL recall without filtering was 0.52 on KiTS and 0.67 on KiRC; Compass filtering also modestly improved FocusMIL.Source: §4.2 · pp. 10–11
  • The unfiltered 3D baseline collapsed to positive predictions, while filtering improved its F1 to 0.70, still below the MIL approaches.The configurations compared 3D ResNet18, ABMIL, and FocusMIL, with and without Compass filtering and, for MIL, Compass pretraining.Source: §4.2 · pp. 9–10
  • F1 = 0.83 was reached by both the strongest FocusMIL configuration and Compass-pretrained, mini-batch ABMIL on TUH.The models reached comparable peak performance on TUH, but differed in out-of-distribution generalization.Source: Table 1 · p. 11, §4.2 · p. 11

5 Discussion

The discussion presents Compass filtering and MIL as complementary: filtering reduces irrelevant input, while slice-based MIL remains stable without filtering. It also identifies dataset and classification challenges that constrain interpretation.

5 Discussion

  • Compass filtering mitigates ABMIL's out-of-distribution failure mode, whereas FocusMIL does not show the same failure, consistent with slice-based MIL's greater stability.The authors suggest filtering limits opportunities to learn spurious correlations from irrelevant scan regions.Source: §5 · p. 12
  • The 3D baseline requires Compass filtering to avoid prediction collapse, while MIL models remain stable on unfiltered inputs.The authors suggest processing individual slices rather than full volumes acts as a regularizing factor in low-data regimes.Source: §5 · p. 12
  • A common error was classifying large benign cysts as malignant, indicating that the task must distinguish malignant from benign abnormalities.The authors suggest adding cyst labels and making the task multi-label as one possible response.Source: §5 · p. 12
  • KiTS and KiRC contain only positive cases, enabling scanner-style shortcuts and limiting external evaluation to sensitivity rather than specificity or precision.The authors note that this makes it difficult to tell whether high external recall reflects generalization or positive-prediction bias.Source: §5 · pp. 12–13

6 Conclusion

The method combines Compass anatomical filtering with multi-instance learning for weakly supervised kidney tumor classification from CT scans. The results support scan-level-label classification, while further work on performance and cross-site generalization is needed before clinical deployment.

6 Conclusion

  • F1 = 0.83 was achieved by the best models on the held-out TUH test set using only scan-level labels.The result demonstrates the feasibility of weakly supervised MIL for kidney tumor classification.Source: §6 · p. 13
  • Compass effectively narrows CT scans to task-specific regions without strong supervision.The approach combines Compass anatomical filtering with multi-instance learning.Source: §6 · p. 13
  • Further work must improve absolute classification performance and cross-site generalization before clinical deployment could be considered.Source: §6 · p. 13
Loading 2609.07178v1…