MetaHOPE: Dual Frameworks in Pathology & MT

• MetaHOPE is a cross-domain term identifying two unrelated artifacts: an image-based MET overexpression predictor and a metaphor translation evaluation framework.
• In computational pathology, MetaHOPE employs a weakly supervised, attention-based multiple instance learning model validated with diverse hold-out and temporal data.
• In machine translation, MetaHOPE adapts a severity-aware HOPE framework to quantify and diagnose metaphor translation errors with refined annotation metrics.

Searching arXiv for MetaHOPE and closely related records to ground the article in the cited preprints. Tool unavailable in this environment. Proceeding with the supplied arXiv metadata and ids (Ingale et al., 2023) and (Liang et al., 1 Jul 2026), which are explicitly provided in the source block. MetaHOPE is the name of two unrelated research artifacts on a weakly supervised, attention-based multiple instance learning model for predicting MET RNA overexpression from hematoxylin and eosin whole-slide images in non-small cell lung adenocarcinoma, introduced in "Prediction of MET Overexpression in Non-Small Cell Lung Adenocarcinomas from Hematoxylin and Eosin Images" (Ingale et al., 2023), and a metaphor-oriented, error severity-aware annotation framework for evaluating machine translation and LLM metaphor translation, introduced in "MetaHOPE: A Metaphor-Oriented Evaluation Framework for Analysing MT and LLM Translation Errors" (Liang et al., 1 Jul 2026). The shared name does not denote a unified methodology. In current usage, it identifies distinct systems in computational pathology and machine translation, so technical discussion requires explicit domain disambiguation.

1. Scope and disambiguation

A common misconception is to treat MetaHOPE as a single framework with transferable assumptions across domains. The available arXiv usage indicates the opposite: the pathology MetaHOPE is a predictive model operating on digitized histology, whereas the translation MetaHOPE is an annotation and evaluation framework operating on metaphor-bearing text segments. This suggests that the name functions as a cross-domain homonym rather than a single research program.

MetaHOPE usage	Domain	Core function
MET RNA overexpression predictor	Computational pathology	Weakly supervised, attention-based predictor from H&E whole-slide images
Metaphor translation evaluation framework	MT and LLM evaluation	Severity-aware annotation of metaphor translation errors

The distinction is substantive. In the pathology setting, MetaHOPE is optimized against a binary label derived from RNA-seq. In the translation setting, MetaHOPE organizes human error annotation, severity scoring, and post-editing workflows. Any comparison between the two therefore operates only at the level of nomenclature, not architecture, objective function, or evaluation protocol.

2. MetaHOPE in computational pathology

In lung adenocarcinoma, the pathology MetaHOPE addresses prediction of MET overexpression from routinely available H&E slides, motivated by the fact that MET protein overexpression is a targetable event in NSCLC and that validated testing such as standardized immunohistochemistry assessment may be inaccessible or consume valuable tissue for a single gene or protein assay (Ingale et al., 2023). The model was trained using a large database of matched H&E slides and RNA expression data to predict MET RNA overexpression directly from H&E images.

The development cohort comprised 1,367 H&E-stained whole-slide images from NSCLC adenocarcinoma patients, all scanned on a Philips UFS scanner. Manual quality control removed cytology specimens, slides with fewer than 100 viable tumor cells, and grossly out-of-focus images; overall fewer than 2% of slides were excluded. To enrich the positive class, cases were subsampled so that approximately half of the development set in cross-validation were MET overexpressors. Independent hold-out evaluation used three datasets: a Philips test set of 589 WSIs, including 300 MET-positive and 289 MET-negative cases; an Aperio test set of 519 WSIs, including 250 MET-positive and 269 MET-negative cases; and a temporal generalization set of 4,331 WSIs collected 6 to 16 months later, stained internally or at external labs, and scanned on both Philips and Aperio systems.

MET overexpression was defined from RNA-seq and copy-number data. RNA-seq quantified MET transcripts in TPM, while DNA copy number variants were called by panel-based NGS and summarized as a segment-length-weighted average copy number. Cases were designated amplified if weighted copy number was at least 5 and wild-type if copy number equaled 2. Comparing these two groups, a $\log_2(\mathrm{TPM}+1)$ threshold of 8 yielded 92% sensitivity and 80% specificity for DNA amplification. The final binary label was

$$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$

For some experiments, a continuous soft label was also defined:

$$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$

This design situates MetaHOPE as an image-based surrogate for a transcriptomic endpoint. A plausible implication is that the model is intended primarily for prescreening and triage rather than as a replacement for confirmatory molecular testing.

3. Learning pipeline for MET overexpression prediction

The preprocessing workflow begins with low-resolution tissue segmentation using a U-Net model following Ronneberger et al. From tissue regions, non-overlapping $224 \times 224$ pixel tiles were extracted at three magnifications: $5\times$, $10\times$, and $20\times$. Tiles overlapping pathologist-drawn annotations such as ink or markers were removed using a second U-Net trained to recognize slide markers. All slides were ICC-profile-mapped to standard sRGB space to reduce inter-scanner and stain variability prior to inference. No additional stain normalization such as Macenko was applied during training, and no geometric augmentations such as flips or rotations were reported. Synthetic color perturbations in brightness, contrast, hue, and saturation were reserved for robustness evaluation.

The model itself is an attention-based multiple instance learning architecture following Ilse et al. A ResNet-18 backbone, pre-initialized on a histopathology classification task, serves as the feature extractor $f_\theta$. Its last convolutional block produces a 512-dimensional embedding $h_i = f_\theta(x_i)$ for each tile $x_i$. Attention pooling then computes an unnormalized attention score using a small multilayer perceptron,

$$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$0

where $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$1 and $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$2, with hidden attention dimension typically $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$3. The normalized attention weights are

$$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$4

These weights define the slide-level representation

$$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$5

and the slide probability

$$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$6

where $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$7, $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$8 is a bias term, and $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$9 is the sigmoid function. Training minimizes slide-level binary cross-entropy,

$$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$0

with L2 weight decay of $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$1 on all parameters.

Model selection used 5-fold patient-level cross-validation with three folds for training, one fold for early stopping or epoch selection, and one fold for validation. Hyperparameters, including learning rate $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$2 and weight decay $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$3, were chosen via grid search on the model-selection fold. Final training reshuffled the development cohort into an approximately 80% training set and 20% model-selection set. Optimization used Adam with constant learning rate $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$4 and weight decay $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$5. Early stopping halted training when validation ROC-AUC on the model-selection set failed to improve for five consecutive epochs, with typical total training lasting 20 to 30 epochs. Each batch consisted of one slide containing all of its tiles, and gradient updates were performed per bag. Best performance was achieved at $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$6 magnification, and the key reproducible setting reported as best performing was tile size $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$7 pixels at $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$8 magnification (Ingale et al., 2023).

4. Validation, explainability, and clinical positioning

On the Philips hold-out test set of 589 WSIs at an operating point targeting 80% sensitivity, the model achieved ROC-AUC $$\mathrm{soft\_label} = \mathrm{sigmoid}\!\left[2 \cdot \left(\log_2(\mathrm{TPM}+1) - 8\right)\right].$$9 with 95th-percentile interval $224 \times 224$0 to $224 \times 224$1, average precision $224 \times 224$2 with 95th-percentile interval $224 \times 224$3 to $224 \times 224$4, sensitivity $224 \times 224$5 with interval $224 \times 224$6 to $224 \times 224$7, specificity $224 \times 224$8 with interval $224 \times 224$9 to $5\times$0, positive predictive value $5\times$1 with interval $5\times$2 to $5\times$3, and $5\times$4 with interval $5\times$5 to $5\times$6 (Ingale et al., 2023). In cross-validation, mean ROC-AUC was $5\times$7 with 95th-percentile interval $5\times$8 to $5\times$9, and the correlation between predicted probability and $10\times$0 was Pearson $10\times$1 with $10\times$2. Magnification mattered: $10\times$3 outperformed $10\times$4 and $10\times$5 with AUC $10\times$6 to $10\times$7 and $10\times$8 and $10\times$9, respectively.

Generalization analyses were extensive. On the Aperio hold-out set, ROC-AUC was $20\times$0 with 95% confidence interval $20\times$1 to $20\times$2, PR-AUC was $20\times$3 with interval $20\times$4 to $20\times$5, sensitivity was 55%, and specificity was 74%. In cross-validation, training only on Philips slides and evaluating on Aperio reduced AUC from $20\times$6 to $20\times$7 with $20\times$8. In the temporal set, collected 6 to 16 months later across internal and external staining sites and both scanners, AUCs remained stable at approximately $20\times$9 for Philips and approximately $f_\theta$0 to $f_\theta$1 for Aperio, with overlapping 95% intervals. Robustness testing applied synthetic perturbations in brightness, contrast, hue, and saturation at ten increasing levels; ROC-AUC and PR-AUC remained within $f_\theta$2 points of unperturbed performance through level 5, and operating-point metrics remained stable through level 5 but degraded thereafter.

Subgroup analyses indicated that large resections achieved AUC $f_\theta$3 whereas core biopsies achieved AUC $f_\theta$4, with $f_\theta$5. By contrast, smoking status, gender, race, and stage showed no statistically significant AUC differences. Specificity to MET was examined by repurposing the predictor to “predict” EGFR, KRAS, BRAF, or TP53 mutations; the resulting AUCs were approximately $f_\theta$6 to $f_\theta$7, with all DeLong comparisons against the MET AUC yielding $f_\theta$8. This supports the interpretation that the learned signal is not a generic oncogenic classifier.

Explainability analyses were centered on attention behavior. The top 10% of tiles by attention score were overlaid on whole-slide images to generate attention heatmaps. A separate tile classifier assigned tiles to six classes—tumor, stroma, epithelium, necrosis, immune, and artifact—and high-attention tiles were $f_\theta$9 tumor versus $h_i = f_\theta(x_i)$0 tumor in the overall tile set, with Wilcoxon $h_i = f_\theta(x_i)$1. UMAP projection of the 512-dimensional ResNet embeddings showed that high-attention tiles co-localized with tumor clusters. Example high-attention regions were densely cellular and pleomorphic tumor areas, whereas low-attention tiles were frequently stroma, necrosis, or artifact.

The stated clinical role is prescreening. A rapid H&E-based predictor could triage cases for confirmatory MET testing by immunohistochemistry or RNA-seq, thereby conserving tissue and laboratory resources. Because the continuous model score correlates with RNA expression, the score could also stratify likely responders once a therapy-defined RNA cutoff is known. The reported limitations are equally explicit: training used a single institution’s data; broader multi-site validation is required; geometric and stain augmentations during training may further improve robustness; and prospective and interventional studies are needed to determine clinical utility and effects on patient care pathways.

5. MetaHOPE in metaphor translation evaluation

In machine translation and natural language processing, MetaHOPE denotes an annotation framework rather than a predictive model. It was proposed as an error severity-aware framework for evaluating metaphor translation because metaphorical expressions exhibit semantic complexity, contextual dependency, and cultural embeddings that can generate ambiguity for NLP models (Liang et al., 1 Jul 2026). General-purpose MT systems may produce overly literal renderings of metaphor-laden multi-word expressions, idioms, or novel metaphors, with consequent loss or distortion of intended meaning. Existing evaluation benchmarks often rely on broad sentence-level metrics such as BLEU, COMET, and MQM or on coarse-grained error taxonomies; the stated motivation for MetaHOPE is that these do not capture the finer nuances of metaphor translation failures.

MetaHOPE adapts the lightweight, severity-based HOPE framework into a metaphor-oriented annotation scheme. Its objectives are to operationalize cognitively motivated concerns in metaphor translation into measurable annotation categories, quantify both frequency and severity of metaphor translation errors, compare GoogleMT, GPT5.4, and Hunyuan-7B on English-Chinese and Chinese-English metaphor translation, and produce human post-edited references and a new parallel resource for future metaphor-sensitive MT research.

The annotation scheme preserves HOPE’s five-level severity taxonomy—minor, medium, major, severe, critical—but replaces the exponential penalty weights $h_i = f_\theta(x_i)$2 with the linear scale $h_i = f_\theta(x_i)$3 to reduce sparsity and inter-annotator disagreement. Errors are categorized into five metaphor-relevant types. IMP denotes structural or emphasis shifts that alter communicative force, such as voice changes or word-order inversion. RAM, or Required Adaptation Missing, denotes failure to render the source metaphor into an appropriate cultural or idiomatic equivalent. MIS denotes loss or distortion of metaphorical meaning through literal rendering. STL denotes attenuation of rhetorical or emotional intensity. PRF denotes fluency or naturalness issues that do not directly affect meaning. Annotators assign each error instance a type $h_i = f_\theta(x_i)$4 and a severity weight $h_i = f_\theta(x_i)$5.

The associated scoring model is additive at the segment level. Let $h_i = f_\theta(x_i)$6 index metaphor-containing segments and $h_i = f_\theta(x_i)$7 the set of errors in segment $h_i = f_\theta(x_i)$8. Then the segment-level penalty is

$h_i = f_\theta(x_i)$9

For a test set of $x_i$0 segments, the average segment penalty is

$x_i$1

The framework also distinguishes full-sentence from metaphor-level penalties:

$x_i$2

The metaphor-error ratio is

$x_i$3

and error-type distributions are

$x_i$4

This formalization makes severity central rather than incidental: error mass, not merely error count, becomes the principal analytic unit.

6. Corpora, empirical behavior, and projected research use

The framework was instantiated using two monolingual, word-level metaphor-annotated corpora. The VU Amsterdam Metaphor Corpus English news subset, annotated via MIPVU, contains 565 metaphor-related words across 200 metaphor-containing sentences. The PSU Chinese Metaphor Corpus news subset, also MIPVU-based, contains 368 metaphor-related words across 200 sentences (Liang et al., 1 Jul 2026). Preprocessing retained only metaphorically tagged nouns, verbs, adjectives, and adverbs; translation was performed at document level to preserve context and then segmented into sentences; and metaphor-related words were aligned manually using semantic rather than strictly lexical alignment so that paraphrase, implicitation, and omission could be captured. For each translation direction, 200 metaphor-bearing sentences were sampled for full test annotation, with a pilot of 20 sentences per direction.

Three systems were evaluated: Google Translate as a proprietary NMT endpoint, GPT5.4 via the OpenAI API using the same few-shot prompt for fairness, and Hunyuan-7B with the Hugging Face default translation prompt. All systems received full-document inputs. Two trained annotators independently applied MetaHOPE guidelines and also produced human post-edited references. Agreement was measured using Krippendorff’s $x_i$5, quadratic-weighted Cohen’s $x_i$6 with severity treated as ordered $x_i$7, Pearson’s $x_i$8, and exact match rate.

On the pilot set, agreement on segment-level summed penalties was highest for GPT5.4, with $x_i$9, weighted $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$00, Pearson $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$01, and exact agreement $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$02. GoogleMT yielded $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$03, weighted $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$04, Pearson $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$05, and exact agreement $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$06. Hunyuan-7B showed the lowest consistency, with $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$07, weighted $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$08, Pearson $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$09, and exact agreement $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$10. The interpretation given is that Hunyuan-7B’s freer paraphrasing and occasional hallucinations increased subjectivity.

Annotator A’s pilot error distributions indicated that GoogleMT accumulated sentence-level penalty $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$11, average $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$12 per sentence, with metaphor-level penalty $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$13 and ratio $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$14; GPT5.4 accumulated $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$15, average $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$16, with $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$17 and ratio $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$18; and Hunyuan accumulated $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$19, average $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$20, with $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$21 and ratio $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$22. By category, GoogleMT had MIS $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$23 and PRF $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$24; GPT5.4 had IMP $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$25, MIS $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$26, and PRF $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$27; Hunyuan had IMP $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$28 and MIS $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$29. The reported interpretation is that more than 90% of GoogleMT and GPT5.4 penalty mass derived from metaphor errors, whereas Hunyuan-7B exhibited a broader error profile, including factual hallucination in addition to metaphor mistranslation.

Qualitative findings differentiated the systems further. GoogleMT and GPT5.4 tended toward preserving source-word order and literalness, described as “rigid tracking,” which often produced Required Adaptation Missing or Mistranslation outcomes. Hunyuan-7B more often generated native-style paraphrases and sometimes better idiomatic equivalents, but at the cost of added or omitted content. The analysis also observed inconsistencies within a single MT system when translating recurring metaphors across sentences, suggesting unstable metaphor handling.

The broader significance of the translation MetaHOPE lies in its attempt to bridge cognitive metaphor theory and empirical MT evaluation. Its five categories correspond to conceptual mapping, cultural remapping, stylistic force, communicative impact, and surface fluency, thereby enabling targeted diagnostics rather than a single monolithic quality score. The stated future directions are scaling to a full test corpus with refined guidelines to improve $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$30 and $$y = 1 \text{ if } \log_2(\mathrm{TPM}+1) \geq 8,\qquad y = 0 \text{ otherwise.}$$31, extending to other domains and languages, investigating semi-automatic or LLM-assisted annotation, and incorporating MetaHOPE into MT training loops or evaluation metrics. A plausible implication is that the framework is designed not only for post hoc analysis but also for future integration into metaphor-sensitive optimization and benchmarking workflows.