Learning to Summarize Radiology Findings

1 Introduction

• Writing impressions is time-consuming and error-prone; omitted crucial findings can cause significant miscommunication.The paper notes that impression generation is highly repetitive with dictating findings.Source: §1 · p. 1
• The task maps radiology findings to impression statements, combining extractive copying of observations with abstractive generation of new conclusions.The customized model encodes study background information and uses it to guide decoding.Source: §1 · pp. 1–2
• The model outperforms non-neural and neural baselines on ROUGE, and a radiologist judged 67% of sampled system summaries at least as good as the human-written references.The authors also report transfer to another organization and occasional summarization of body parts unseen during training.Source: §1 · p. 2
• The paper reports this as the first attempt at automatic summarization of radiology findings into natural-language impressions.The authors describe their work as the first attempt in this direction.Source: §1 · p. 2

2 Related Work

• Early summarization systems primarily selected input sentences, using topic-word scores, classifiers, latent semantic analysis, or graph-based ranking.Graph methods represent sentences as vertices and similarities as edges before scoring and selecting sentences.Source: §2 · p. 2
• Neural summarization enables abstractive generation, with later systems adding copying, coverage, reinforcement learning, or sentence selection followed by rewriting.Pointer-generators copy input elements to handle out-of-vocabulary words; coverage addresses repetition.Source: §2 · p. 2
• Prior radiology-report work focused mainly on classification and information extraction; this paper presents the first attempt to summarize findings into natural-language impressions.Earlier work included named-entity extraction and identification of asserted or negated disease entities.Source: §2 · pp. 2–3

3 Task Definition

• Given findings tokens x of length N, the task seeks a summary token sequence y of arbitrary length L that captures salient and clinically significant findings.The passage formalizes findings as x = {x1, x2, . . . , xN} and the summary as y = {y1, y2, . . . , yL}.Source: §3 · p. 3
• The formulation leaves the summary length arbitrary rather than fixing it in advance.The formal definition describes L as an arbitrary summary length.Source: §3 · p. 3
• The mapping from findings x to summary y may be modeled unsupervised or learned from a dataset of findings-summary pairs.Source: §3 · p. 3

4.1 Neural Sequence-to-Sequence Model

• A bidirectional LSTM encodes findings, and a separate LSTM decoder generates the output sequence one step at a time.The initial decoder state is set to the encoder's final combined state; attention uses a weighted sum of input states at each decoding step.Source: §4.1 · p. 3
• At each decoding step, attention scores input states using the decoder state and an input hidden state.The resulting attention distribution weights input states to provide salient input information for decoding.Source: §4.1 · p. 3

4.2 Pointer-Generator Network

• The pointer-generator network can either generate a vocabulary word or copy a word directly from the findings at each decoding step.Its copy distribution reuses the attention distribution over the input.Source: §4.2 · p. 3

4.3 Incorporating Study Background Information

• A separate attentional encoder represents the study background, and its representation guides every state in the decoder.The background encoder produces a weighted representation of salient background information, which modifies the decoder LSTM recurrent kernel.Source: §4.3 · p. 4

5.1 Data Collection

• The dataset contains 87,127 reports from the top 12 PACS body parts and was randomly split into 70% training (60,990), 10% development (8,712), and 20% test (17,425) sets.Reports were collected from 2000 to 2014. Common body parts with more than 10,000 reports were subsampled to 10,000.Source: §5.1 · p. 5

5.2 Baseline Models

• Baselines include LSA- and graph-based extractive systems and pointer-generator models using findings alone or background prepended to findings.The extractive systems select high-scoring sentences; the pointer-generator coverage mechanism was omitted because it did not improve summary quality in this task.Source: §5.2 · p. 5

5.3 Experimental Setup

• Models are evaluated with ROUGE-1, ROUGE-2, and ROUGE-L F1 scores for unigram, bigram, and longest-common-sequence overlap.Source: §5.3 · p. 5
• 100-dimensional GloVe vectors trained on 4.5 million radiology reports initialize neural-model embeddings and improved performance by about 1 ROUGE-L score.Source: §5.3 · p. 5
• Prepending background can lead extractive baselines to select background sentences as summaries, deteriorating performance.Source: §5.3 · p. 5
• Neural models use teacher forcing, negative log-likelihood with Adam, two-layer bidirectional LSTM encoders, and beam search of size 5 at inference.The authors tuned hyperparameters on the development set and capped output length at 100 tokens.Source: §5.3 · p. 6

6.1 Main Results

• Our background-guided model achieves the best scores on all ROUGE metrics, while prepending background to the pointer-generator hurts performance.The pointer-generator substantially outperforms the non-neural extractive baselines on all metrics.Source: §6.1 · p. 6

6.2 Clinical Validity with Radiologist Evaluation

• 67 of 100 sampled summaries were judged at least as good as human-written summaries by a board-certified radiologist.The radiologist judged 51 pairs equivalent, preferred the system summary in 16, and preferred the human-written summary in 33.Source: §6.2 · pp. 6–7, Table 2 · p. 7

6.3 Does the model transfer to reports from another organization?

• On 2,691 Indiana University reports, the model substantially outperformed the extractive baseline, producing grammatical and clinically meaningful summaries.The best model trained on the authors’ dataset was run directly on the Indiana University test set.Source: Table 3 · p. 7, §6.3 · p. 7

6.4 Does the model transfer to body parts unseen during training?

• With chest or abdomen reports withheld from training, summaries degraded substantially; knee summaries degraded less, presumably because related musculoskeletal reports were available.For an unseen knee study, the model included salient observations such as tissue swelling and fracture while copying “knee” from the findings.Source: Table 4 · p. 8, §6.4 · p. 8

6.5 What is the model missing on?

• 63% of sampled development examples were qualitatively close to the reference, while missing critical information was the most common error at 24%.The model tended to omit follow-up procedures recommended by radiologists because they were often absent from findings and required significant study understanding and domain knowledge.Source: Table 5 · p. 8, §6.5 · p. 9

• The paper proposes generating radiology impressions from findings with neural sequence-to-sequence learning.To the authors’ knowledge, this was the first attempt in this direction.Source: §7 · p. 9
• A customized model uses encoded background information to guide the decoding process.The dataset was collected from actual hospital studies.Source: §7 · p. 9
• The model outperforms non-neural and neural baselines and generates summaries with clinical validity and cross-organization transferability.Source: §7 · p. 9