💻 computer science

YOLO-PVC: 2D-to-3D Consolidation of Slice-wise Detections for Volumetric Liver Tumor Localization in MRI

The paper proposes YOLO-PVC, a lightweight and model-agnostic framework that consolidates fragmented 2D slice-wise detections into stable 3D volumetric liver tumor localizations in MRI by enforcing depth continuity and refining axial extent through robust statistics and a calibration module, achieving superior performance over existing aggregation baselines.

Original authors: Talha Waqas, Mounir Lahlouh, Kawther Taibouni, Mahnoor Waqas, Salar Ahmed, Sébastien Mulé, Yasmina Leroul-Chenoune

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Talha Waqas, Mounir Lahlouh, Kawther Taibouni, Mahnoor Waqas, Salar Ahmed, Sébastien Mulé, Yasmina Leroul-Chenoune

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a 3D model of a hidden treasure inside a giant, transparent block of Jell-O. You can't see the whole block at once, so you have to look at it one thin slice at a time, like flipping through the pages of a very thick book. In the world of medical imaging, doctors do something similar with MRI scans. They take a 3D image of a patient's liver and slice it up into hundreds of 2D pictures. The goal is to find tumors—lumps of bad cells—and draw a perfect 3D box around them to measure their size and plan treatment.

For a long time, computers were great at looking at these 2D slices one by one, finding the tumors, and drawing boxes on the flat pages. But when you try to stack those flat boxes back up to make a 3D shape, things get messy. It's like trying to build a tower out of Jell-O slices where some slices are missing, some have extra blobs of jelly that aren't part of the tower, and the edges don't line up perfectly. The result is a wobbly, broken tower that doesn't look like the real thing. This is a huge problem because if the computer can't measure the tumor accurately in 3D, it's hard for doctors to know if a treatment is working or how big the surgery needs to be. Scientists have been trying to figure out how to glue these 2D slices together into a solid, reliable 3D shape without needing super-expensive, slow computers that can look at the whole block at once.

This is where a new idea called YOLO-PVC comes in. Think of it as a clever "glue and ruler" system that takes those messy, wobbly 2D slices and snaps them together into a sturdy 3D box. The researchers behind this method realized that the biggest mistake other computers make is just stacking the slices on top of each other like a pile of pancakes. If one pancake is slightly bigger or smaller than the one below it, the whole tower gets crooked.

Instead of just stacking, YOLO-PVC acts like a smart detective. First, it looks at the stack of slices and says, "Okay, these three slices are clearly part of the tumor, but this one in the middle is just a glitch—ignore it." It filters out the noise and finds the longest, most continuous line of slices where the tumor actually exists. Then, instead of taking the absolute biggest or smallest edge it sees (which might be a mistake), it uses a "percentile" trick. Imagine you have a hundred measurements of a tumor's width; instead of picking the biggest one (which might be a fluke) or the average (which might be too small), it picks a spot that covers 90% of the real data but ignores the top 10% of weird outliers. This creates a box that is tight but doesn't accidentally cut off the edges of the tumor.

But the researchers didn't stop there. They noticed that even with this smart gluing, the box was still a little too short from top to bottom, like a shoe that fits the foot but misses the toes. To fix this, they added a tiny, super-fast "calculator" (a small neural network) that learns from past mistakes. It looks at the shape of the slices and says, "Hey, based on how this tumor looks, we need to stretch the box up a tiny bit more." This little adjustment makes the final 3D box fit the real tumor much better.

The paper shows that this method works really well. When they tested it on liver MRI scans with three different types of tumors, their new method managed to fit the 3D box around the tumor with an accuracy score (called IoU3D) of 0.710. That's a big jump compared to older methods, which only scored around 0.424 to 0.596. It also fixed the problem of the box being too short, reducing the error from missing about 4.5 slices of depth to almost zero.

The best part is that this doesn't require a super-computer. It works on the same fast, lightweight tools that doctors already use to look at 2D slices, just adding a smart step to put them together. The researchers found that this approach is so good at handling missing slices or weird errors that even if you accidentally delete 10% of the slices, the system still builds a decent box. It's a practical, efficient way to turn a messy stack of 2D pictures into a reliable 3D map, helping doctors see the full picture of a liver tumor without needing to build a whole new, heavy computer system from scratch.

Technical Summary: YOLO-PVC

Problem Statement
Volumetric object detection in medical imaging, particularly for liver tumor localization in MRI, is critical for diagnosis and treatment planning. While fully 3D convolutional networks offer robust inter-slice context, they require dense volumetric annotations and substantial GPU memory, limiting their scalability in resource-constrained clinical settings. Conversely, modern 2D object detectors (e.g., YOLO) are computationally efficient and scalable but, when applied slice-wise to volumetric data, produce independent predictions along the depth axis. Naïve stacking of these 2D predictions often results in fragmented 3D bounding boxes, inconsistent axial extents, and sensitivity to false positives or missing slices, leading to anatomically implausible volumetric estimates. The central challenge addressed is how to recover reliable, anatomically consistent 3D bounding boxes from noisy slice-wise detections without resorting to heavy 3D architectures.

Methodology: YOLO-PVC
The authors propose YOLO-PVC (Percentile-based Volumetric Consolidation), a lightweight, model-agnostic framework designed to consolidate slice-wise 2D detections into a coherent 3D bounding box. The pipeline operates on 3D bounding boxes derived from expert radiologist segmentation (used as ground truth) but does not require voxel-level supervision during training or inference. The method consists of four key stages:

Depth Continuity Filtering: To address isolated false positives, the method identifies the longest approximately contiguous subset of slices containing detections. A gap tolerance parameter ( $g$ ) allows for limited missed slices caused by detector instability while suppressing isolated outliers.
Percentile-Based Spatial Aggregation (PVC): Instead of using simple averaging or Min-Max stacking, the framework aggregates planar coordinates ( $x, y$ ) using empirical percentiles (e.g., 10th and 90th percentiles). This approach treats slice-wise detections as samples from a noisy geometric process, providing controlled robustness against extreme values while preserving tumor coverage.
Depth Padding: To counter systematic underestimation of axial extent near object boundaries, a symmetric padding parameter ( $\delta$ ) is applied to the initial depth bounds ( $z_{min}, z_{max}$ ) derived from the filtered slice indices.
MLP-Based Calibration (MLP-PVC): A lightweight Multi-Layer Perceptron (MLP) module further refines the axial extent. This module takes an 8D feature vector per slice (derived from bounding box geometry, confidence, and relative position) and predicts per-slice depth offsets. These offsets are aggregated via mean pooling to correct residual depth bias, shifting the estimation toward near-zero error without altering lateral localization.

Key Contributions

Structured Estimation Framework: The paper reframes slice-to-volume consolidation from a deterministic post-processing step into a structured estimation problem that explicitly models outliers and missing detections.
Robust Aggregation: The introduction of percentile-based statistics for lateral fusion and continuity-aware filtering offers a significant improvement over traditional heuristics like confidence-weighted averaging or median fusion.
Lightweight Calibration: The MLP-based calibration module introduces negligible computational overhead (approx. 2.7K parameters) while effectively correcting systematic depth bias, a common failure mode in 2D-to-3D pipelines.
Annotation Efficiency: The framework operates purely on bounding-box predictions, making it suitable for clinical environments where dense voxel-level annotations are unavailable or too costly to acquire.

Experimental Results
Experiments were conducted on a retrospective dataset of 142 arterial-phase liver MRI volumes containing three tumor categories: Hepatocellular Carcinoma (HCC), Intrahepatic Cholangiocarcinoma (CCA), and Mixed tumors. The framework was evaluated against five baseline aggregation strategies (Confidence-Weighted Average, Trimmed Mean, Median Fusion, Continuity-Aware Linking, and Min-Max Stacking).

Performance: The heuristic PVC variant achieved an overall IoU3D of 0.665, outperforming the best baseline (Min-Max Stacking) which scored 0.596. The calibrated Hybrid MLP-PVC variant further improved performance to an IoU3D of 0.710, representing a +19.2% relative improvement over Min-Max Stacking.
Depth Accuracy: A critical finding was the reduction of axial bias ( $\Delta z$ ). Baseline methods exhibited severe underestimation ( $\Delta z \approx -4.51$ ). The heuristic PVC reduced this to $-2.61$, and the Hybrid MLP-PVC brought it to near-zero ($-0.125$).
Planar Stability: The method maintained high planar overlap, with BEV IoU scores around 0.78, indicating that depth refinement did not compromise lateral localization.
Robustness: Under stress tests involving random slice removal (10% missing) and the injection of adversarial outlier slices, the framework demonstrated graceful degradation, maintaining geometric coherence, though performance dropped significantly under extreme structured outlier conditions.

Significance and Claims
The paper claims that structured geometric consolidation provides an effective and practical solution for volumetric liver tumor localization in clinical MRI. The authors emphasize that the primary constraint in slice-to-volume reconstruction is axial instability rather than planar localization error. By enforcing inter-slice continuity and utilizing robust statistical aggregation, YOLO-PVC achieves consistent improvements in volumetric overlap and reduced depth bias.

The significance lies in offering a modular and annotation-efficient alternative to 3D segmentation networks. It enables the use of efficient 2D detectors for 3D tasks in settings where only bounding-box labels are available and GPU memory is limited. The results suggest that minimal learned depth reasoning, combined with robust heuristics, can substantially enhance volumetric reconstruction quality across heterogeneous tumor morphologies without the computational cost of full 3D architectures.

Technical Summary: YOLO-PVC

More like this