PaperlayerView exact revision on arXiv

Source-linked AI summary

EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction

Michelle Espranita Liman, Ioana Ionac, Alexander Müller, Eimo Martens, Daniel Rueckert, Philip Müller

arXiv:2610.13599v1cs.LG

TL;DR

Existing echo representation methods do not explicitly model measurements as an intermediate representation between views and diagnosis. EchoTab learns factorized, measurement-specific tokens through tabular reconstruction; experiments report improved diagnosis performance on two datasets and attention aligned with established measurement-view relationships.

  • Problem

    Existing echo representation methods do not explicitly represent individual measurements as an intermediate representation that downstream diagnoses can selectively use.

    Source: §1 · pp. 1–2

  • Method

    EchoTab uses measurement-specific tokens that attend to relevant echo views and are trained by tabular reconstruction to predict measurement values.

    Source: §1 · p. 2

  • Results

    EchoTab improves diagnosis performance on two datasets over pre-training methods that learn generic study-level representations, and its learned attention aligns with established measurement-view relationships.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Measurement-grounded tokens provide a factorized, more interpretable basis for diagnosis than direct mapping from echo views to diagnosis.

    Source: §6 · p. 9

    ECHOTAB requires specifying diagnosis-relevant measurements and evaluates only averaging as the strategy for combining their tokens.

    Source: §6 · p. 10

Abstract

from arXiv · show

Diagnostic information in multi-view medical imaging studies, like echocardiography, is distributed across views, each capturing different aspects of cardiac structure and function. In clinical practice, these views are first interpreted through quantitative measurements, which then support specific diagnoses. However, existing echo representation learning methods do not explicitly model measurements as an intermediate representation between views and diagnosis. As a result, individual measurements are not explicitly represented in a form that downstream diagnostic tasks can selectively use, hindering accurate diagnosis. Inspired by this hierarchical clinical workflow, we introduce EchoTab, a measurement-grounded representation learning framework that represents multi-view echocardiography studies using measurement-specific tokens as an intermediate representation between echo views and diagnosis. Each measurement token selectively attends to the views relevant to its estimation and is trained through tabular reconstruction to predict the corresponding measurement value. The resulting tokens form a factorized echo representation, which encode clinically meaningful structural and functional properties and can be selectively used for downstream diagnosis. Unlike the commonly used contrastive approaches, EchoTab does not require negative samples or large batch sizes. Experiments on two datasets demonstrate that EchoTab learns effective measurement-grounded representations that improve diagnosis classification in multi-view echocardiography and its learned attention aligns with established measurement-view relationships used by cardiologists. Our code is accessible at https://github.com/michelleespranita/EchoTab.

1 INTRODUCTION

Echo interpretation proceeds from views to quantitative measurements to diagnosis, but existing echo representation methods do not explicitly model measurements as a factorized intermediate representation. EchoTab learns measurement-specific tokens through tabular reconstruction and uses them for diagnosis-specific selection, improving diagnosis performance on two datasets over generic study-level pre-training methods.

1 INTRODUCTION

  • Existing echo methods directly predict diagnoses, require finetuning, or use report retrieval, while measurement-aware approaches do not explicitly represent individual measurements as continuous values.Measurements provide continuous, dense supervision, whereas categorical diagnosis labels lose information about phenotype strength; for example, left ventricular ejection fraction is derived from A2C and A4C views and informs systolic dysfunction severity.Source: §1 · pp. 1–2
  • EchoTab assigns each measurement a token that selectively attends to relevant echo views, forming a factorized intermediate representation between views and diagnosis.A prediction head reconstructs the corresponding measurements from the tokens, encouraging them to capture clinically meaningful structural and functional properties.Source: §1 · p. 2
  • EchoTab improves diagnosis performance on two datasets over pre-training methods that learn generic study-level representations, without requiring negative samples or large batch sizes.Its learned attention over echo views is consistent with established measurement-view relationships used by cardiologists.Source: §1 · p. 2

2 ECHOTAB

ECHOTAB encodes each echo view, then uses measurement-specific tokens to gather relevant view information and reconstruct measurements. The resulting factorized tokens support downstream diagnosis, while the reconstruction objective accounts for sparse measurements with differing scales.

2 ECHOTAB

  • ECHOTAB encodes each view, forms measurement-specific tokens that gather relevant view information, reconstructs measurements, and supports diagnosis using relevant tokens.The framework is trained end-to-end with tabular reconstruction before relevant measurement tokens are selected for downstream diagnosis.Source: §2 · p. 2

2.1 ECHO ENCODER

  • Each study contributes one selected clip per available view; the encoder processes views independently and retains local patch tokens with learnable view embeddings.Background tokens below the stated intensity threshold are masked; the passage reports that this improves view-classification performance.Source: §2.1 · pp. 2–3

2.2 TABULAR DECODER

  • Measurement queries cross-attend to echo tokens to produce independent measurement-factorized representations, without self-attention between measurement tokens.The decoder is designed to aggregate information from the views relevant to each measurement.Source: §2.2 · pp. 3–4

2.3 PREDICTION HEAD

  • The prediction head reconstructs each measurement from its token, with independent, shared, or partially shared parameterizations.Independent heads scale in parameters with the number of measurements, shared heads do not, and partially shared heads use hypernetwork-generated measurement-specific low-rank corrections.Source: §2.3 · pp. 4–5

2.4 TABULAR RECONSTRUCTION OBJECTIVE

  • ECHOTAB trains with masked, variance-normalized mean squared error over observed study-measurement pairs, scaling errors by each measurement’s training-set variance.Variance normalization puts measurements on a comparable footing so measurements with large absolute values do not dominate the loss.Source: §2.4 · p. 5

3 EXPERIMENTS

ECHOTAB is evaluated on two echocardiography datasets for diagnosis classification and for alignment between its measurement-view attention and cardiologist-selected views. It improves average diagnosis AUC over comparison methods and matches cardiologist views in most evaluated measurement-view pairs.

3.1 DATA

  • The evaluation uses MIMIC-IV-Echo with 4,514 patients and 7,144 studies, and an internal cohort with 7,404 patients and 9,431 studies.Source: §3.1 · p. 5

3.2 BASELINES AND EVALUATION STRATEGY

  • Diagnosis evaluation uses k-nearest neighbors with k = 5 on study-level embeddings, comparing ECHOTAB with echo foundation models and measurement-based contrastive and MLP baselines.The MLP baseline uses independent measurement-prediction heads and ECHOTAB's tabular reconstruction objective.Source: §3.2 · pp. 5–6
  • For each diagnosis, ECHOTAB averages its relevant measurement tokens, while baselines without measurement-token factorization use averaged view representations.The study also tests diagnosis-relevant views as a baseline prior and restricts ECHOTAB to relevant views.Source: §3.2 · p. 6

3.3 DIAGNOSIS CLASSIFICATION

  • ECHOTAB achieves the highest average AUC among echo foundation models, contrastive baselines, and the MLP baseline: 0.748 on MIMIC and 0.778 on the internal dataset.It ranks first or second on 8 of 9 MIMIC diagnoses and 7 of 8 internal diagnoses.Source: §3.3 · p. 6, Table 1 · p. 7
  • On MIMIC, ECHOTAB exceeds view-level SupCon by 4.4 average AUC points (0.748 vs. 0.704) while using 1 GPU rather than 8 GPUs.On the internal dataset, it beats the strongest baseline on 7 of 8 diagnoses and raises average AUC by 1.7 points (0.778 vs. 0.761).Source: §3.3 · p. 6

3.4 ATTENTION MAP ANALYSIS

  • ECHOTAB matches cardiologist-selected views in 16 of 20 measurement-view pairs (80%), with deviations described as anatomically plausible.The three measurement disagreements involve biplane LVEDV and LVEF and TAPSE.Source: §3.4 · pp. 6–7, Figure 3 · p. 7

4 ABLATIONS

The ablations identify the partially shared prediction head and diagnosis-relevant measurement-token selection as important design choices for ECHOTAB’s diagnosis performance. They also show how the decoder and query/head configuration affect the resulting representations.

4.1 ARCHITECTURAL ABLATION

  • With independent heads, adding the tabular decoder improves diagnosis classification over the MLP baseline on MIMIC (0.697 vs. 0.720) and the internal cohort (0.761 vs. 0.766).The authors interpret this as evidence that querying views through the decoder yields a more informative study representation than averaging view embeddings.Source: §4.1 · p. 8
  • Independent prediction heads with measurement queries slightly underperform a single query, while a shared head restores alignment and improves classification to 0.732 on MIMIC and 0.773 internally.Without measurement-specific queries, a shared head gives the same representation for every measurement and prevents measurement-specific predictions.Source: §4.1 · p. 8
  • The partially shared head yields the best diagnosis classification in both cohorts: 0.748 on MIMIC and 0.778 on the internal dataset.It balances cross-measurement alignment with measurement-specific flexibility and achieves the best MIMIC measurement prediction (0.418), while essentially tying the fully shared head (0.447) and MLP (0.446) on internal measurement prediction (0.445).Source: §4.1 · p. 8, Table 2 · p. 8

4.2 MEASUREMENT TOKENS FOR DIAGNOSIS CLASSIFICATION

  • Averaging all measurement tokens or using incorrect tokens reduces classification performance, indicating that selecting diagnosis-relevant tokens is crucial for accurate diagnosis.Source: §4.2 · pp. 8–9

5 RELATED WORK

Prior echo and vision-tabular approaches do not explicitly use measurement representations as an intermediate step from views to diagnosis. ECHOTAB combines measurement concepts with query-based decoding and trains them through tabular reconstruction rather than contrastive objectives.

Echo foundation models

  • Echo foundation models use masked prediction, indirect measurement exposure through reports, or parallel measurement and diagnosis prediction rather than measurement representations as an intermediate diagnostic step.ECHOTAB explicitly uses measurement representations between echo views and diagnosis.Source: §5 · p. 9

Vision-tabular learning

  • Vision-tabular methods MMCL and TGV use contrastive objectives that generally benefit from larger, more diverse negative-sample sets and larger batches, unlike ECHOTAB’s tabular reconstruction objective.Source: §5 · p. 9

Concept bottlenecks and query-based decoders

  • ECHOTAB combines concept-bottleneck-style measurement tokens with query-based decoding that reads information from multi-view echoes through cross-attention.Source: §5 · p. 9

6 DISCUSSION AND LIMITATIONS

ECHOTAB uses diagnosis-relevant measurement tokens to factorize echo studies, improving diagnostic accuracy over generic study-level representations. Its limitations include specified diagnosis–measurement mappings and a single tested token-aggregation strategy.

6 DISCUSSION AND LIMITATIONS

  • Diagnosis-relevant measurement tokens improve diagnostic accuracy over generic study-level representations and provide a more interpretable basis for diagnosis.The framework is presented as applicable in principle to multi-view imaging with associated measurements beyond echocardiography.Source: §6 · p. 9
  • The framework requires specifying diagnosis–measurement mappings, introducing domain knowledge, and tests only averaging as a way to combine relevant tokens.The authors identify learning these relationships and exploring more sophisticated aggregation as future work.Source: §6 · p. 10

AI USE STATEMENT

  • The authors report using generative AI for methodology feedback, implementation, result interpretation, literature identification, and drafting or editing, with human review and responsibility for the final work.They state that AI tools were not used to generate synthetic datasets or develop theoretical models or hypotheses.Source: AI USE STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The code is provided publicly, while MIMIC-IV-Echo access requires PhysioNet credentialing; processing, measurements, hyperparameters, and baseline details are documented in appendices.The cited code repository is https://github.com/michelleespranita/EchoTab.Source: REPRODUCIBILITY STATEMENT · p. 10

A DATA PROCESSING

The data pipeline builds paired echo–measurement datasets, filters measurements and studies, harmonizes values, and controls video sampling and patient-level splits.

A DATA PROCESSING

  • The internal cohort and MIMIC-IV-Echo were processed independently for measurement selection and study filtering, with patient-level train, validation, and test splits.No patient appears in more than one split; Table 5 reports the splits as patient and study counts.Source: A DATA PROCESSING · p. 13, Table 5 · p. 14
  • Measurement filtering removes redundant variants and body-size-indexed values in the internal cohort, merges synonymous MIMIC fields, and retains M-mode measurements in both datasets.Table 4 summarizes the number of measurements remaining after each filtering step; dashes mark steps that do not apply.Source: A DATA PROCESSING · p. 13, Table 4 · p. 13
  • Study filtering excludes records without retained measurements or videos from the specified view set, leaving 9,430 internal-cohort studies and 7,063 MIMIC studies.Source: A DATA PROCESSING · p. 13
  • Measurement values below the training-set first percentile or above its 99th percentile are set to missing, and units are harmonized across datasets.Examples include converting milliseconds to seconds and millimeters to centimeters.Source: A DATA PROCESSING · p. 13
  • Training randomly samples one available video per view type at each epoch, while evaluation and k-NN classification use the first available video for deterministic inputs.Source: A DATA PROCESSING · p. 13

B ADDITIONAL RESULTS ON DIAGNOSIS CLASSIFICATION

When baselines receive diagnosis-relevant views as a diagnosis-specific prior, ECHOTAB retains the highest average AUC on both datasets. Its advantage spans most diagnoses in each dataset.

B.1 CAN A DIAGNOSIS-SPECIFIC PRIOR FOR THE BASELINES CLOSE THE GAP TO ECHOTAB?

  • The comparison gives ECHOTAB all views plus diagnosis-relevant measurement tokens, while restricting each baseline to its diagnosis-relevant views.This setup tests whether an analogous diagnosis-specific prior closes the performance gap.Source: §B.1 · p. 14
  • ECHOTAB exceeds the strongest baseline by 3.1 average AUC points on MIMIC (0.748 vs. 0.717 for MLP) and 3.0 on the internal dataset (0.778 vs. 0.748 for study-level SupCon).Baselines were re-evaluated using diagnosis-relevant views as an analogous diagnosis-specific prior.Source: §B.1 · p. 14
  • ECHOTAB has the best average AUC on 6 of 9 MIMIC diagnoses and 7 of 8 internal-dataset diagnoses.Source: §B.1 · p. 14

B.2 DOES ECHOTAB BENEFIT FROM DIAGNOSIS-RELEVANT VIEWS AS PRIOR?

Restricting ECHOTAB to diagnosis-relevant views lowers average AUC on both datasets. The authors report that using all views avoids per-diagnosis view curation.

B.2 DOES ECHOTAB BENEFIT FROM DIAGNOSIS-RELEVANT VIEWS AS PRIOR?

  • The authors suggest the decoder may weight relevant views more effectively or extract complementary information from other views, allowing ECHOTAB to use all study views without per-diagnosis view curation.Source: §B.2 · p. 14
  • Restricting ECHOTAB to diagnosis-relevant views lowers average AUC from 0.748 to 0.725 on MIMIC and from 0.778 to 0.753 on the internal dataset.Table 7 reports average AUC and indicates that the decoder benefits from access to all views.Source: Table 7 · p. 14, §B.2 · p. 14

C DATASET DETAILS OF DIAGNOSIS CLASSIFICATION TASKS

The section specifies the diagnosis-task data and experimental setup, including label handling, echo-token selection, and model baselines. It also reports that ECHOTAB outperforms baselines under diagnosis-specific prior information.

C DATASET DETAILS OF DIAGNOSIS CLASSIFICATION TASKS

  • The internal cohort uses four severity levels, while MIMIC uses seven, including trace and intermediate grades; studies lacking a task label are excluded.Consequently, task totals can differ within a split.Source: C DATASET DETAILS OF DIAGNOSIS… · pp. 14–15
  • Table 6 reports that ECHOTAB outperforms baselines when both receive diagnosis-specific prior information, using all views with diagnosis-relevant measurement tokens.The baselines and reference foundation models use only diagnosis-relevant views.Source: Table 6 · p. 15
  • Training classes are balanced as much as possible, while validation and test retain original distributions; cardiologist-defined mappings link diagnoses to measurements and views.These mappings identify diagnosis-relevant measurements and views.Source: C DATASET DETAILS OF DIAGNOSIS… · p. 15

D.3 ECHO TOKEN SEQUENCE

  • Local EchoPrime patch tokens with background patches removed yield the best k-NN view-classification performance among the tested configurations.This configuration is used for EchoPrime-based models so measurement queries can learn to attend to the correct views.Source: §D.3 · p. 16, Table 12 · p. 18

F HYPERPARAMETER SETTINGS

  • ECHOTAB trains for 30 epochs with AdamW, using two cross-attention blocks in its tabular decoder and two-layer MLP prediction heads.The echo encoder learning rate differs by dataset; the partially shared head uses rank-8 low-rank updates.Source: F HYPERPARAMETER SETTINGS · p. 17

G HYPERNETWORK VARIANTS

  • Hypernetwork variants L, M, and S trade fewer parameters for more shared-body passes, with M used by default.Their factor-generation strategies reduce output dimensions from 12,288 (L) to 6,144 (M) and 768 (S), while S requires 16 passes.Source: G HYPERNETWORK VARIANTS · p. 18, Table 13 · p. 18

E MEASUREMENTS

  • The MIMIC and internal cohorts use 45 and 67 measurements, respectively, and Figure 4 depicts echo-study distributions by view and split.Videos are cropped and resized, sampled to 16 frames, and zero-padded when short.Source: §D.2 · pp. 15–16, Figure 4 · p. 18, Figure 5 · p. 19

H.1 MLP

  • The MLP baseline averages study-level view embeddings and applies a separate measurement-specific head, following ECHOTAB’s prediction-head architecture.The section also identifies MLP, contrastive, and supervised-contrastive baselines using the same EchoPrime-initialized encoder and view set.Source: §H.1 · p. 19, H BASELINES · p. 19

H.2 CONTRASTIVE

The contrastive baselines align echo and measurement representations, with study- and view-level sampling variants. SupCon instead defines positives by measurement similarity, guaranteeing at least one positive per anchor.

H.2 CONTRASTIVE

  • The contrastive baseline uses a CLIP-style objective to align paired echo and measurement representations, with symmetric losses for both directions.The projections are 128-dimensional, and the temperature is set to 0.07.Source: §H.2 · pp. 19–20
  • Study-level contrastive learning pools a study’s views, while view-level learning pairs each view with its study’s measurements and provides more samples and negative pairs.Measurement representations come from a frozen, precomputed TARTE encoder.Source: §H.2 · pp. 20–21

H.3 SUPERVISED CONTRASTIVE (SUPCON)

  • The SupCon baseline uses normalized measurements to define tabular similarity and selects positives relative to each anchor rather than by a fixed threshold.All remaining batch samples serve as negatives, and the selection guarantees at least one positive per anchor.Source: §H.3 · p. 21
  • SupCon also uses study- and view-level variants, with batch sizes of 32 and 192, respectively.Both variants train for 30 epochs; temperature is 0.07 and the positive-selection margin h is 0.05.Source: §H.3 · p. 21

I DETAILED RESULTS ON MEASUREMENT PREDICTION

The section presents ECHOTAB measurement-prediction R2 results on MIMIC at both per-measurement and per-section levels.

I DETAILED RESULTS ON MEASUREMENT PREDICTION

Loading 2610.13599v1…