Generating specific sections of the report such as Impression and Recommendations. 

Report corrector and prompt insertion.

Provide clinical guidance for breast reporting.

Personalizing reports to match the physician's individual style and preference. Supporting clear, professional, and accurate reporting.

Report Impression 1.0

Report Recommendation 1.0

All outputs are reviewable suggestions — the reading physician retains full control.

After completing the Findings section.

To help generate clear summary sections.

Not a diagnostic tool — it does not analyze images. 

Not a full report generator. 

Not autonomous — all outputs require physician review. 

Not a replacement for the reading physician’s judgment and editing.

Reading physicians can face inconsistency and variability in impression statements due to time constraints, fatigue, or differences in dictation style.

The Impression Agent converts report findings into consistent, concise, and clinically accurate list of impression statements.

Gathering all the recommendations in the report can be time-consuming for the reading physicians.

The Recommendation Agent extracts any mention of recommendations from the report and generates a list of concise recommendations that is organized under a new section in the report.

Report Assistant 3.0

Report Impression 1.0

Report Recommendation 1.0

Report Assistant takes a report as the input and makes parallel call to the two sub-agents to process it.​

Report Assistant then takes the output and parses the result in a JSON structured output with the list of Impressions and Recommendations. If there are no recommendations, an empty list is sent out.

1. Exams with a report.

2. Exams with an Impression section​. 

3. Gender: Male or female.

4. Acceptable facilities and shortlisted physicians who generate high quality reports.

5. Patient age: More or equal to 21.

Cosine Similarity​

Jaccard Similarity​

Levenshtein Similarity

True Positives (TP)​

True Negatives (TN)​

False Positives (FP)​

False Negatives (FN)​

Accuracy​

Precision​

Sensitivity​

Specificity​

F1 Score

1. Similarity Score​

2. Hallucination Score​

3. Correctness Score​

GT: “No acute intracranial abnormality is seen.”​

AI: “There is no evidence of acute intracranial pathology.”​​

GT: “Mild disc degeneration at L4-L5.”​

AI: “Fracture seen in the right femur.”​

GT: “No acute intracranial abnormality is seen.”​

AI: “There is no evidence of acute intracranial pathology.”​​

GT: “Mild disc degeneration at L4-L5.”​

AI: “Fracture seen in the right femur.”​

GT: “No acute intracranial abnormality is seen.”​

AI: “There is no evidence of acute intracranial pathology.”​​

GT: “Mild disc degeneration at L4-L5.”​

AI: “Fracture seen in the right femur.”​

Helps assess over generation or under generation.​

Indicates whether the AI is being too verbose or missing expected findings.​

Important for evaluating clinical usefulness and readability.

No acute intracranial abnormality.​

Mild chronic microvascular ischemic changes.

No evidence of acute intracranial pathology.​

Mild chronic ischemic changes noted.​

Age-related brain atrophy present.​

No mass effect or midline shift.

TP

At least one IMPRESSION item​

At least one IMPRESSION item​

TN

No IMPRESSION item

No IMPRESSION item

FP

At least one IMPRESSION item

No IMPRESSION item

FN

No IMPRESSION item

At least one IMPRESSION item​

TP

At least one RECOMMENDATION item

At least one RECOMMENDATION item

TN

No RECOMMENDATION item

No RECOMMENDATION item

FP

At least one RECOMMENDATION item

No RECOMMENDATION item

FN

No RECOMMENDATION item

At least one RECOMMENDATION item

Accuracy​

Proportion of total correct predictions.​

Precision

How many predicted positives are correct.​

Sensitivity​

How many actual positives were correctly predicted.​

Specificity​

How many actual negatives were correctly predicted.​

F1 Score​

Harmonic mean of precision and sensitivity.​

How closely the AI output matches the Ground Truth in structure and meaning.

Assesses whether the AI introduces unsupported or fabricated content.​

Compares the AI output to the Original Report.​

Penalizes clinically incorrect or invented statements.

Measures how accurately the AI reflects the clinical content in the Ground Truth.

Evaluates how closely the AI-generated section aligns in meaning and structure with the Ground Truth (GT). This score focuses on semantic overlap, organization, and phrasing consistency.

GT: "Slightly coarsened hepatic echotexture. 3.6 humeri simple right renal cyst."​

AI Output: "Slightly coarsened liver echotexture. 3.6 cm right renal simple cyst."

GT: "No renal stones or acute process seen."​

AI Output: "Lumbar fusion with mild S-shaped scoliosis and multilevel degenerative joint disease (DJD). Vascular calcifications noted."

Assesses whether the AI introduces unsupported or fabricated findings not present in the Original Report. It focuses on clinical safety and factual grounding.

Original Report: "LIVER: Slightly coarsened echotexture without focal hepatic mass. Midclavicular length: 13.9 cm. RIGHT KIDNEY: Size: 10.7 x 4.0 x 5.2 cm (L x W x H). 3.6 cm right renal simple cyst. " ​

AI Output: "Slightly coarsened liver echotexture. 3.6 cm right renal simple cyst."

Original Report: "PRE-FINDINGS: EXAM: CT OF THE HEAD/BRAIN WITHOUT INTRAVENOUS CONTRAST COMPARISON: None. INDICATIONS: TRAUMA TECHNIQUE: An axial spiral CT examination was performed without intravenous contrast. In addition, sagittal and coronal reformatted images were created and reviewed. This CT uses dose modulation, interval reconstruction, and/or weight-based dosing when appropriate to reduce radiation dose to as low as reasonably achievable. FINDINGS: BRAIN INCLUDING MENINGES AND EXTRA AXIAL SPACES: Acute abnormality. Severe chronic small vessel disease. Generalized cerebral volume loss. The hippocampi appear decreased in volume, with extra vacuum enlargement of the temporal horns. No upward bowing of the corpus callosum or asymmetric narrowing of the hy- convexity sulci is suggests hydrocephalus. VASCULATURE: No atherosclerotic changes in the major vessels. PITUITARY: No mass or deformity of the sella. SKULL: No significant lesion or fracture. SINUSES: Limited views demonstrate no significant mucosal thickening or fluid. ORBITS: Left-sided scleral band. Right lens replacement. OTHER: There are scattered foci of air in the left masticator space and in the left preauricular region. Trace left mastoid effusion."​

AI Output: "Acute abnormality in the brain. Severe chronic small vessel disease. Generalized cerebral volume loss with decreased hippocampal volume and enlargement of temporal horns. Left-sided scleral band and right lens replacement. Scattered foci of air in the left masticator space and left preauricular region. Trace left mastoid effusion. ”

Evaluates how accurately the AI captures the clinical meaning of the Ground Truth. Focuses on factual consistency and correct medical interpretation.

GT: "Slightly coarsened hepatic echotexture. 3.6 humeri simple right renal cyst."​

AI Output: "Slightly coarsened liver echotexture. 3.6 cm right renal simple cyst."

GT: "No renal stones or acute process seen."​

AI Output: "Lumbar fusion with mild S-shaped scoliosis and multilevel degenerative joint disease (DJD). Vascular calcifications noted."

5 → Fully Correct​

4 → Minor inaccuracies​

3 → Half correct items​

2 → Major inaccuracies​

1 → Fully incorrect

No Hallucination → All AI results are grounded and are based on the original report​

Not Applicable → There is no recommendation in the original report and AI correctly generated no recommendations; Therefore Hallucination score doesn't apply here.​

Contains Hallucination → The generated results contain information outside the original report

5-Fully Correct​

266 (80.6%)​

4-Minor Inaccuracies​

49 (14.8%)​

3-Half Correct Items​

12 (3.6%)​

2-Major Inaccuracies

3 (0.9%)​

1-Fully Incorrect​

0 (0.0%)​

Count

330​

Mean

4.75

Median

5​

Std Dev

0.56​

Min

2

Max

5

No Hallucination​

321 (97.3%)​​

Contains Hallucination​

9 (2.7%)​

5-Fully Correct​

266 (80.6%)​

4-Minor Inaccuracies​

49 (14.8%)​

3-Half Correct Items​

12 (3.6%)​

2-Major Inaccuracies

3 (0.9%)​

1-Fully Incorrect​

0 (0.0%)​

Count

330​

Mean

4.75

Median

5​

Std Dev

0.56​

Min

2

Max

5

No Hallucination​

321 (97.3%)​​

Contains Hallucination​

9 (2.7%)​

5-Fully Correct​

306 (92.7%)​

4-Minor Inaccuracies​

6 (1.8%)​

3-Half Correct Items​

2 (0.6%)​

2-Major Inaccuracies

2 (0.6%)​

1-Fully Incorrect​

14 (4.2%)​

Count

330​

Mean

4.78​

Median

5​

Std Dev

0.85​

Min

1

Max

5

No Hallucination​

103 (31.2%)​

Not Applicable​

227 (68.8%)​