Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1865
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-18T22:02:30.844207Z
status: published
lens: g_space
sector: health
headline: Multi-task deep learning integrating pretreatment MRI and whole slide images predicts induction chemotherapy response and survival in locally advanced nasopharyngeal carcinoma
dek: Predicting response to induction chemotherapy (IC) and overall survival (OS) is critical for optimizing treatment in patients with locally advanced nasopharyngeal carcinoma (LANPC). This study aimed to develop and validate a multi-task deep learning model integrating pretreatment MRI and whole slide images (WSIs) to predict IC response and OS in LANPC. Pretreatment MRI and WSIs from 404 patients with LANPC were retrospectively collected to construct a multi-task model (MoEMIL) for the simultaneous prediction of…
gain_title: A model combining pretreatment MRI and pathology slides separated good from poor induction chemotherapy responders and outperformed TNM staging on both response and survival prediction in a test set.
problem_title: (none)
trace_subject: (none)
gain_reading: A model combining pretreatment MRI and pathology slides separated good from poor induction chemotherapy responders and outperformed TNM staging on both response and survival prediction in a test set.
gain_evidence: MoEMIL effectively stratified patients into good and poor IC response groups | In the test set, MoEMIL significantly outperformed the deep learning radiomics model and TNM staging for both endpoints.
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers built a multi-task deep learning system called MoEMIL that reads pretreatment MRI scans and whole slide pathology images from patients with locally advanced nasopharyngeal carcinoma. It predicts early response to induction chemotherapy and overall survival at the same time.

The work points toward earlier, more individualized treatment planning, because response and prognosis could be estimated before therapy begins. The authors describe it only as a possible adjunct tool, and the evidence comes from retrospective data, so its value in actual clinical decisions has not been shown.
limitation: The cohort had no progressive disease cases, so the poor responder group consisted only of stable disease patients, and the data were retrospective; the improvement over the pathomics model for IC response was only borderline.
tag: Evidence-backed gain
key_points: The study used pretreatment MRI and whole slide images from 404 retrospectively collected patients with locally advanced nasopharyngeal carcinoma. | Whole slide images were processed with multi-instance learning, while MRI features came from PyRadiomics and a ResNet50 network, fused through a multi-gate mixture-of-experts design. | Areas under the curve for IC response were 0.917, 0.869, and 0.801 in the train, validation, and test sets. | The model split patients into high- and low-risk overall survival groups (P < 0.05).
rundown: The model, MoEMIL, takes two data types: radiology features from MRI and histopathology features from digitized tissue slides. A multi-gate mixture-of-experts architecture fuses them so one network handles both response and survival prediction.

In the test set it significantly outperformed a deep learning radiomics model and TNM staging on both endpoints. Against a pathomics-only model, it showed a significant gain for overall survival (P = 0.0049) and a borderline gain for IC response (P = 0.064).

For interpretation, the team used clustering-constrained attention multiple instance learning and gradient-weighted class activation mapping to visualize which image regions drove predictions.
sources:
- peer_reviewed | PLOS Digital Health | https://doi.org/10.1371/journal.pdig.0001378 | 2026-10-05
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

8e5c4187013381bb7b4242d9354ec60b83526e47152c6d396c38456dcc47727d

previous

0000000000000000000000000000000000000000000000000000000000000000