@inproceedings{salme-etal-2026-wrote,
title = "{I} Wrote This Because of That: Interpretable Radiology Report Generation via Visual Prototypes",
author = "Salm{\`e}, Marco and
Siciliano, Federico and
Silvestri, Fabrizio and
Soda, Paolo and
Sicilia, Rosa and
Guarrasi, Valerio",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1708/",
doi = "10.18653/2026.findings-emnlp.1708",
pages = "36511--36527",
ISBN = "979-8-89176-462-0",
abstract = "Automated Radiology Report Generation (RRG) could substantially reduce documentation workload by supporting physicians in routine clinical practice. Clinical adoption, however, requires each generated finding to be supported by inspectable evidence that a physician can verify. Existing systems either provide only post-hoc explanations weakly coupled to the prediction, or require costly region-level supervision to achieve spatial grounding. To address this gap, we introduce ProtoRRG, an RRG framework that grounds generation in visual prototypes. Specifically, each generated finding is tied to a dual form of evidence: localised image exemplars learned without anatomical supervision, and retrieved prior reports that supply case-based linguistic precedents. A multi-task training strategy reconciles the competing representational demands of prototype-based prediction and autoregressive decoding within a shared architecture. On MIMIC-CXR, ProtoRRG improves state-of-the-art clinical F1 and recall by 2.9{\%} and 10{\%}, respectively, with gains also demonstrated on IU X-Ray. An evaluation against anatomical annotations confirms that prototype activations localise within the correct regions, indicating spatially faithful evidence for the generated findings. These results demonstrate that combining visual prototypes with case-based retrieval improves both diagnostic accuracy and report quality while preserving the inspectable evidence that clinical deployment demands."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="salme-etal-2026-wrote">
<titleInfo>
<title>I Wrote This Because of That: Interpretable Radiology Report Generation via Visual Prototypes</title>
</titleInfo>
<name type="personal">
<namePart type="given">Marco</namePart>
<namePart type="family">Salmè</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Federico</namePart>
<namePart type="family">Siciliano</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fabrizio</namePart>
<namePart type="family">Silvestri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paolo</namePart>
<namePart type="family">Soda</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rosa</namePart>
<namePart type="family">Sicilia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Valerio</namePart>
<namePart type="family">Guarrasi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Automated Radiology Report Generation (RRG) could substantially reduce documentation workload by supporting physicians in routine clinical practice. Clinical adoption, however, requires each generated finding to be supported by inspectable evidence that a physician can verify. Existing systems either provide only post-hoc explanations weakly coupled to the prediction, or require costly region-level supervision to achieve spatial grounding. To address this gap, we introduce ProtoRRG, an RRG framework that grounds generation in visual prototypes. Specifically, each generated finding is tied to a dual form of evidence: localised image exemplars learned without anatomical supervision, and retrieved prior reports that supply case-based linguistic precedents. A multi-task training strategy reconciles the competing representational demands of prototype-based prediction and autoregressive decoding within a shared architecture. On MIMIC-CXR, ProtoRRG improves state-of-the-art clinical F1 and recall by 2.9% and 10%, respectively, with gains also demonstrated on IU X-Ray. An evaluation against anatomical annotations confirms that prototype activations localise within the correct regions, indicating spatially faithful evidence for the generated findings. These results demonstrate that combining visual prototypes with case-based retrieval improves both diagnostic accuracy and report quality while preserving the inspectable evidence that clinical deployment demands.</abstract>
<identifier type="citekey">salme-etal-2026-wrote</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1708</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1708/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>36511</start>
<end>36527</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T I Wrote This Because of That: Interpretable Radiology Report Generation via Visual Prototypes
%A Salmè, Marco
%A Siciliano, Federico
%A Silvestri, Fabrizio
%A Soda, Paolo
%A Sicilia, Rosa
%A Guarrasi, Valerio
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F salme-etal-2026-wrote
%X Automated Radiology Report Generation (RRG) could substantially reduce documentation workload by supporting physicians in routine clinical practice. Clinical adoption, however, requires each generated finding to be supported by inspectable evidence that a physician can verify. Existing systems either provide only post-hoc explanations weakly coupled to the prediction, or require costly region-level supervision to achieve spatial grounding. To address this gap, we introduce ProtoRRG, an RRG framework that grounds generation in visual prototypes. Specifically, each generated finding is tied to a dual form of evidence: localised image exemplars learned without anatomical supervision, and retrieved prior reports that supply case-based linguistic precedents. A multi-task training strategy reconciles the competing representational demands of prototype-based prediction and autoregressive decoding within a shared architecture. On MIMIC-CXR, ProtoRRG improves state-of-the-art clinical F1 and recall by 2.9% and 10%, respectively, with gains also demonstrated on IU X-Ray. An evaluation against anatomical annotations confirms that prototype activations localise within the correct regions, indicating spatially faithful evidence for the generated findings. These results demonstrate that combining visual prototypes with case-based retrieval improves both diagnostic accuracy and report quality while preserving the inspectable evidence that clinical deployment demands.
%R 10.18653/2026.findings-emnlp.1708
%U https://aclanthology.org/2026.findings-emnlp.1708/
%U https://doi.org/10.18653/2026.findings-emnlp.1708
%P 36511-36527
Markdown (Informal)
[I Wrote This Because of That: Interpretable Radiology Report Generation via Visual Prototypes](https://aclanthology.org/2026.findings-emnlp.1708/) (Salmè et al., Findings 2026)
ACL