pith. sign in

arxiv: 2605.05891 · v1 · submitted 2026-05-07 · 💻 cs.CV · cs.AI· cs.LG

MTL-MAD: Multi-Task Learners are Effective Medical Anomaly Detectors

Bogdan Alexandru Bercean , Florinel Alin Croitoru , Vlad Hondru , Ciprian Mihai Ceausescu , Andreea Iuliana Ionescu , Radu Tudor Ionescu This is my paper

Pith reviewed 2026-05-08 14:29 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords anomalymodelmedicalmulti-tasktaskbmadduringeffective
0
0 comments X

The pith

A multi-task learner using Mixture-of-Experts on multiple self-supervised tasks from scratch outperforms state-of-the-art methods for anomaly detection across medical image modalities on the BMAD benchmark.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Detecting unusual patterns in medical scans is hard because training data usually contains only normal examples. Instead of relying on one auxiliary task like image reconstruction, this method trains a single model on several tasks at once using a Mixture-of-Experts design, where different expert networks specialize in different tasks. When a new scan arrives, the model’s difficulty in solving those tasks serves as an anomaly score. The authors test this on BMAD, a benchmark covering many scan types, and report better results than existing techniques while also producing visual maps that highlight suspicious regions.

Core claim

The empirical results indicate that our multi-task learner is an effective anomaly detector, outperforming all state-of-the-art competitors on BMAD.

Load-bearing premise

By carefully integrating multiple proxy tasks, the joint model effectively learns a robust representation of normal anatomical structures, so that anomaly scores can be derived based on how well the multi-task learner solves each task during inference.

read the original abstract

Anomaly detection in medical images is a challenging task, since anomalies are not typically available during training. Recent methods leverage a single pretext task coupled with a large-scale pre-trained model to reach state-of-the-art performance. Instead, we propose to learn multiple self-supervised and pseudo-labeling tasks from scratch, using a joint model based on Mixture-of-Experts (MoE). By carefully integrating multiple proxy tasks, the joint model effectively learns a robust representation of normal anatomical structures, so that anomaly scores can be derived based on how well the multi-task learner (MTL) solves each task during inference. We perform comprehensive experiments on BMAD, a recent benchmark that comprises a broad range of medical image modalities. The empirical results indicate that our multi-task learner is an effective anomaly detector, outperforming all state-of-the-art competitors on BMAD. Moreover, our model produces interpretable anomaly maps, potentially helping physicians in providing more accurate diagnoses.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on the domain assumption that multiple proxy tasks jointly produce a robust normal representation; no free parameters, invented entities, or additional axioms are specified.

axioms (1)
  • domain assumption Multiple self-supervised and pseudo-labeling tasks can be jointly optimized to capture normal anatomical structures better than single-task alternatives.
    Invoked to justify the multi-task MoE design and the inference-time anomaly scoring rule.

pith-pipeline@v0.9.0 · 5497 in / 1176 out tokens · 28488 ms · 2026-05-08T14:29:47.679351+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.