PaperlayerView exact revision on arXiv

Source-linked AI summary

An interpretable deep learning method for early hepatocellular carcinoma diagnosis on multiphase CT

Emma Sarfati, Alexandre Bône, Pietro Gori, Sisi Yang, Marc-Michel Rohé, François Nicolas, Jeong-Min Lee, Jeong Hee Yoon, Maxime Ronot, Isabelle Bloch, Christophe Aubé

arXiv:2610.29616v1cs.CVcs.AI

TL;DR

Small liver lesions can be difficult to characterize consistently on multiphase CT, while prior deep-learning approaches often provide limited insight into their reasoning. This paper proposes dLI-RADS, combining neural-network predictions and handcrafted LI-RADS-related measurements for lesion-level HCC risk estimation. It was non-inferior to LI-RADS on internal accuracy and external sensitivity, with the authors noting a need for prospective workflow evaluation.

  • Problem

    Assessing LI-RADS major features on CT is challenging and subject to inter-reader variability, particularly for small lesions, while prior deep-learning work often focused on larger lesions and offered limited explainability.

    Source: Introduction · p. 3

  • Method

    dLI-RADS estimates lesion-level HCC risk from manually segmented multiphase CT by combining neural-network predictions and handcrafted LI-RADS-related features through logistic regression.

    Source: Introduction · p. 3, The dLI-RADS method · pp. 5–6

  • Results

    dLI-RADS was non-inferior to LI-RADS for internal HCC diagnostic accuracy (80% vs. 74%, p = 0.04) and external sensitivity (93% vs. 78%, p < 0.001).

    Source: Discussion · p. 15

  • Takeaways & Limitations

    Combined dLI-RADS and LI-RADS risk categories achieved very high specificity, suggesting possible complementarity that requires independent external validation.

    Source: Discussion · p. 16

    Manual lesion segmentation limits practical use in routine clinical workflows, and the study could not quantify the effect of inter-reader segmentation variability.

    Source: Discussion · p. 16

Abstract

from arXiv · show

Early diagnosis of hepatocellular carcinoma is critical to preserve access to curative treatment options. The Liver Imaging Reporting and Data System (LI-RADS) provides a standardized framework for non-invasive diagnosis on multiphase computed tomography, based on major visual imaging features. However, these features may be difficult to assess consistently, particularly in small liver lesions. We developed dLI-RADS, an interpretable artificial intelligence method designed to estimate the probability of hepatocellular carcinoma from multiphase computed tomography images of manually segmented liver lesions, while remaining aligned with LI-RADS-related imaging concepts. dLI-RADS combines neural network-derived predictions with handcrafted quantitative imaging measurements reflecting LI-RADS major features and integrates these inputs using logistic regression. In this multicenter retrospective study, patients with multiphase computed tomography and at least one liver lesion were included, while lesions larger than 40 mm were excluded. Histology served as the reference standard for hepatocellular carcinoma, and a composite reference standard was used for non-hepatocellular carcinoma lesions. The lesion-level accuracy of dLI-RADS and LI-RADS were compared on the internal evaluation set using 10,000-repetition bootstrap non-inferiority tests, with 5% margins and significance levels. On the external test set, sensitivities were similarly compared. dLI-RADS was non-inferior to LI-RADS for lesion-level accuracy in the internal evaluation set, achieving 80% accuracy compared with 74% for LI-RADS (p = 0.04)...

1 Guerbet Research, Villepinte, France

Non-invasive HCC diagnosis on CT remains challenging, especially for small lesions where visual features are difficult to assess consistently. dLI-RADS was proposed to combine deep and handcrafted features for early-stage lesion diagnosis while balancing performance and explainability.

Introduction

  • LI-RADS standardizes multiphase imaging interpretation, but assessing major features on CT remains challenging and subject to inter-reader variability.The difficulty is particularly relevant for small lesions.Source: Introduction · p. 3
  • Prior radiomics mined predefined handcrafted-feature libraries, while deep-learning studies often focused on larger lesions and offered limited insight into model reasoning.The cited discussion identifies particular interest in diagnosing smaller early-stage nodules.Source: Introduction · p. 3
  • dLI-RADS combines deep and handcrafted features to diagnose early-stage HCC from manually segmented multiphase CT lesions, aiming to balance clinical performance and explainability.Its performance was compared with LI-RADS in two independent European and Asian cohorts using non-inferiority tests.Source: Introduction · p. 3

Methods

The retrospective multicenter study used separate European and Asian cohorts, with a patient-level internal split and an independent external evaluation set. dLI-RADS processes tri-phasic CT patches and lesion masks through handcrafted and deep feature extraction, followed by logistic regression.

Study Design and Datasets

  • The main European cohort enrolled 442 patients with compensated cirrhosis, while the secondary Asian cohort included 513 patients with de novo liver nodules.The main cohort used histology with a composite standard for non-malignant cases; the secondary cohort used histology for malignant tumors and at least six-month follow-up for benign lesions.Source: Study Design and Datasets · p. 4
  • The main cohort was split 70/30 into training and internal evaluation sets with patient-level segregation; the non-overlapping secondary cohort served as external evaluation.Exclusions included incomplete multiphase CT, lesions larger than 40 mm, missing definitive diagnosis, and LR-1, LR-2, or LR-M classification.Source: Study Design and Datasets · p. 4

Ethics declarations

  • The study used informed consent in the main cohort and institutional review board approval with waived consent in the secondary cohort; all data were de-identified.The main cohort was registered as #NCT00848952, and the study reported adherence to relevant guidelines and the Declaration of Helsinki.Source: Ethics declarations · p. 4

The dLI-RADS method

  • dLI-RADS extracts eight intermediate features from tri-phasic CT patches centered on pre-segmented lesions, then combines them with logistic regression to estimate HCC probability.The features comprise handcrafted measurements related to LI-RADS major features and lesion size, plus neural-network probabilities for APHE, EC, NPW, and HCC presence.Source: The dLI-RADS method · pp. 5–6, Figure 1 · p. 5

Image pre-processing

  • CT phases were co-registered, resampled to 2.00 × 0.76 × 0.76 mm3, and stacked with the lesion segmentation in 24 × 96 × 96-voxel patches.Scans were clipped between -100 and 400 Hounsfield units.Source: Image pre-processing · p. 6

Training and inference

The neural networks generate deep features that are combined with handcrafted descriptors in a logistic regression model. Thresholds are selected on the training set to align operating points with radiologists’ sensitivity and specificity.

Training and inference

  • Four neural networks shared an architecture with five convolutional blocks (5 × 5 × 5 kernels, 2 × 2 × 2 max pooling), two dense layers and softmax output; they trained for 300 epochs, batch size 32, using AdamW.The learning rate was 10-5 and weight decay was 10-3.Source: Training and inference · p. 6
  • Deep-learning features from the training images were concatenated with handcrafted descriptors to fit logistic regression, whose raw output was a continuous HCC probability between 0 and 1.Operating thresholds were selected to minimize the summed absolute distance from radiologists’ sensitivity and specificity on the training set.Source: Training and inference · p. 6
  • The two selected thresholds were named dLI-RADS-4,5 and dLI-RADS-5 to correspond directly to the LI-RADS-4,5 and LI-RADS-5 operating points.The selection strategy was related to index-of-union threshold selection and adapted to established clinical operating points.Source: Training and inference · p. 7

Statistical analysis

The study compared dLI-RADS and LI-RADS using bootstrap non-inferiority tests, and assessed complementary risk groupings and feature importance. The evaluation sets comprised patients and liver nodules with distinct characteristics across training, internal, and external cohorts.

Statistical analysis

  • Performance was assessed with sensitivity, specificity, and accuracy for LI-RADS-5 and combined LI-RADS-4,5 groups; external LI-RADS comparison was limited to sensitivity.Bootstrap confidence intervals used 10,000 repetitions, with non-inferiority tests using a 5-percentage-point margin and 5% significance level.Source: Statistical analysis · p. 7
  • Compound risk groups were formed by summing dLI-RADS and LI-RADS categories, and feature importance was explored by evaluating feature combinations on the training set.The best-performing combination for each number of retained features was then evaluated on the internal set.Source: Statistical analysis · p. 7

Patient and lesion characteristics

  • The inclusion flow reduced 442 initially recruited main-cohort patients to 205 included patients, while 133 of 513 initially recruited patients entered the secondary cohort.The flowchart covered exclusions including incomplete multiphase CT, lesions larger than 40 mm, missing definitive diagnosis, and LI-RADS inapplicability.Source: Figure 2 · p. 8, Patient and lesion characteristics · p. 10
  • The 338 patients contributed 143 training, 62 internal, and 133 external cases; the corresponding lesion counts were 196, 84, and 171.Mean lesion sizes were 18 mm in training and internal evaluation and 22 mm externally; the external set had 89% cirrhosis, compared with all patients in the training and internal sets.Source: Patient and lesion characteristics · p. 10

Qualitative examples

Figure 3 pairs representative CT patches with radar plots to illustrate how dLI-RADS and LI-RADS assess liver lesions, including concordant and discordant classifications.

Qualitative examples

  • Figure 3 pairs CT patches with radar plots summarizing derived and radiological features, illustrating concordant and discordant dLI-RADS and LI-RADS assessments.The examples include HCC lesions, a hemangioma, and a metastasis.Source: Figure 3 · p. 10, Qualitative examples · p. 11
  • For one HCC, dLI-RADS captured subtle arterial-phase hyper-enhancement that the radiologist did not include in LI-RADS, while both methods captured visible washout.The radiologist did not judge the arterial-phase hyper-enhancement sufficient for the LI-RADS assessment.Source: Qualitative examples · p. 11
  • Two central examples received opposed risk groups (5 vs. 3 and 3 vs. 5), while a difficult HCC confounder was classified dLR-5 and LR-5.The confounder showed APHE, NPW, and a mildly enhancing capsule.Source: Qualitative examples · p. 11

Diagnostic performance of dLI-RADS and LI-RADS

On internal evaluation, dLI-RADS was non-inferior to LI-RADS for accuracy at the maximal-risk threshold. External sensitivities and an exploratory combined assessment are also reported.

Diagnostic performance of dLI-RADS and LI-RADS

  • 80% vs. 74% accuracy for dLR-5 versus LR-5 on internal evaluation was non-inferior (p = 0.04); sensitivities were 84% versus 78% (p = 0.03).The non-inferiority margin was 5%.Source: Diagnostic performance of dLI-RADS and… · p. 12, Table 2 · p. 12
  • For internal high-risk groups, dLR-4,5 versus LR-4,5 achieved 90% versus 88% sensitivity, 42% versus 46% specificity, and 75% dLI-RADS accuracy.The reported p-values were 0.10 for sensitivity, 0.44 for specificity, and 0.16 for accuracy.Source: Diagnostic performance of dLI-RADS and… · p. 12
  • External sensitivities were 93% versus 78% for dLR-5 versus LR-5 and 97% versus 94% for dLR-4,5 versus LR-4,5; dLI-RADS specificity/accuracy were 74%/86% and 62%/84%, respectively.LI-RADS specificity and accuracy could not be computed externally because only HCC nodules were characterized.Source: Diagnostic performance of dLI-RADS and… · pp. 12–13
  • dLI-RADS AUC was 76% on internal and 84% on external evaluation; Figure 4 marks its dLR operating points and LI-RADS comparators when available.External LI-RADS operating points are shown where only sensitivity could be derived.Source: Figure 4 · p. 11, Diagnostic performance of dLI-RADS and… · p. 13

Combined model-based and radiological assessment

Feature importance

Ablations found the complete eight-feature dLI-RADS model performed best, while HCC-DL was consistently selected in reduced configurations.

Feature importance

  • Table 3 compares configurations reduced from eight to one intermediate feature, selecting the best-performing combination on training data for internal evaluation.Source: Table 3 · p. 14
  • 75.6% AUC was highest for the complete eight-feature dLI-RADS model; HCC-DL alone achieved 70.6% AUC and was selected consistently across reduced configurations.The ablation table reports AUC for train-selected feature combinations tested on internal evaluation.Source: Feature importance · p. 14, Table 3 · p. 14
  • Adding handcrafted and deep LI-RADS-inspired features improved performance over the HCC-DL feature alone.The findings support combining data-driven representations with clinically structured descriptors.Source: Feature importance · pp. 14–15

Discussion

dLI-RADS was non-inferior to LI-RADS on the internal accuracy and external sensitivity comparisons, with 86% standalone external accuracy. The discussion highlights its small-lesion focus and interpretable feature framework while noting limits in interpretability, segmentation, and external comparison.

Discussion

  • 80% vs. 74% accuracy internally and 93% vs. 78% sensitivity externally supported dLI-RADS non-inferiority to LI-RADS on those respective endpoints.The internal comparison had p = 0.04; the external comparison had p < 0.001.Source: Discussion · p. 15
  • The combined maximal-risk category reached 96% specificity at 69% sensitivity, but this exploratory analysis was internal-only and requires external confirmation.The authors suggest possible complementarity between model-based and radiological assessment.Source: Discussion · pp. 15–16
  • The study focused on small lesions, averaging 18 mm internally and 22 mm externally, and used histology plus composite or follow-up reference standards across cohorts.The authors argue that small-lesion evaluation is relevant because radiologists are generally confident diagnosing lesions larger than 40 mm unaided.Source: Discussion · p. 15
  • Six of eight intermediate features were interpretable as LI-RADS major-feature probabilities, allowing radiologists to compare model-inferred features with their own assessments.The authors distinguish this two-step approach from end-to-end models that offer little insight into decision-making.Source: Discussion · pp. 15–16
  • Interpretability remains partial because a black-box feature played a major role in the final risk score.The authors describe the two-step pipeline as a compromise between performance and explainability.Source: Discussion · p. 16
  • Manual lesion segmentation limits routine practicability, and the study could not quantify segmentation variability or fully compare external accuracy because LI-RADS features were unavailable for non-HCC lesions.The authors call for end-to-end segmentation studies and prospective evaluation; validation data came from 17 centers in Europe and Asia.Source: Discussion · p. 16

Competing interests

Several authors are employed by Guerbet; the remaining authors report no competing interests.

Competing interests

  • Emma Sarfati, Alexandre Bône, Sisi Yang, Marc-Michel Rohé, and François Nicolas are employed by Guerbet; the other authors declare no competing interests.Source: Competing interests · p. 18

Tables

The supplied table captions describe patient characteristics, diagnostic-performance comparisons, and feature-reduction ablations.

Tables

  • Table 1 presents clinical characteristics for the training, internal evaluation, and external evaluation sets.Source: Table 1 · p. 28
  • Table 2 covers sensitivity, specificity, accuracy, and AUC for dLI-RADS and LI-RADS risk groups, with bootstrap non-inferiority tests using 5% margins and significance levels.Its caption cautions that reported p-values test non-inferiority, not superiority.Source: Table 2 · p. 29
  • Table 3 describes ablations reducing the number of dLI-RADS intermediate features from eight to one, selecting configurations on the training set and testing them internally.Source: Table 3 · p. 30
Loading 2610.29616v1…