Experience-Guided Self-Adaptive Cascaded Agents for Breast Cancer Screening and Diagnosis with Reduced Biopsy Referrals

The paper proposes BUSD-Agent, an experience-guided self-adaptive cascaded multi-agent framework for breast ultrasound screening and diagnosis that leverages a memory bank of historical decision trajectories to dynamically adjust escalation thresholds, significantly reducing unnecessary biopsy referrals and improving specificity without requiring model parameter updates.

The Gist

Imagine you are running a very busy, high-stakes airport security checkpoint for breast health. Every day, thousands of people (ultrasound images) walk through. The goal is to catch the dangerous "smugglers" (cancer) while letting the innocent travelers (benign lumps or normal tissue) pass through without hassle.

The Problem:
In the current system, the security guards are too paranoid. They flag almost everyone for a full-body scan and a pat-down (a biopsy). This causes long lines, huge stress for the passengers, and wastes the time of the elite special agents (radiologists) who should only be dealing with the real threats.

The Solution: BUSD-Agent
The authors of this paper built a smart, two-stage security system called BUSD-Agent. Think of it as a team of two different types of security officers working together, guided by a "memory book" of past cases.

Stage 1: The "Quick-Check" Officer (The Screening Agent)

• Who they are: A fast, lightweight officer who uses a set of standard metal detectors and scanners (AI classification models).
• What they do: They look at the ultrasound image and give a quick "thumbs up" or "thumbs down."
• The Old Way: If the scanner beeped even slightly, the officer would send the passenger to the next stage for a full search. This caused too many false alarms.
• The New Way (The Magic): This officer doesn't just rely on the scanner. Before making a decision, they flip open a Memory Book.
  • The Analogy: Imagine the officer sees a passenger with a weird-looking belt buckle. Instead of panicking, they check their memory book: "Wait, I saw 10 people with similar buckles last week. They all turned out to be harmless fashion accessories. I trust that pattern."
  • Because of this memory, the officer confidently says, "You're clear, go on," for many people who would have been stopped before. This stops the traffic jam.

Stage 2: The "Special Ops" Team (The Diagnostic Agent)

• Who they are: A highly trained, elite team of experts with microscopes, X-rays, and detailed maps (advanced tools like lesion segmentation and radiological description).
• What they do: They only see the people the Quick-Check officer wasn't sure about. They do a deep dive to decide if a biopsy (a tissue sample) is actually needed.
• The New Way: Just like the first officer, this team also checks the Memory Book. They look for cases that look exactly like the current one.
  • The Analogy: If the current case looks like a "suspicious" lump, the team checks the book: "In the past, when we saw this exact shape and texture, it turned out to be cancer 9 times out of 10. Let's send this one for a biopsy."
  • Or, they might see a pattern that usually turns out to be harmless and decide to skip the invasive test.

The Secret Sauce: The "Memory Book" (Experience-Guided Adaptation)

The most brilliant part of this system is how it learns without retraining.

• Old AI: To learn, you usually have to feed it thousands of new examples and reprogram its brain (update parameters). This is slow and expensive.
• BUSD-Agent: It uses In-Context Learning. It doesn't change its brain; it just changes its reference material.
  • Every time a patient gets a biopsy and the result comes back (Pathology), that result is written into the Memory Book as a "Decision Trajectory."
  • When a new patient arrives, the system instantly finds the most similar past cases in the book and says, "Hey, look at how we handled these guys before. Let's do the same thing."

The Results: Why This Matters

By using this "experience-guided" approach, the system achieved amazing results across 10 different datasets:

1. Fewer False Alarms: It stopped sending innocent people to the "Special Ops" team. The number of unnecessary escalations dropped from 85% to 59%.
2. Fewer Biopsies: Because fewer innocent people were escalated, the number of unnecessary, painful, and expensive biopsies dropped from 60% to 37%.
3. Still Safe: Crucially, they didn't miss any real cancer. The system remained just as good at catching the real threats (Sensitivity) while becoming much better at ignoring the harmless ones (Specificity).

In a Nutshell

Think of BUSD-Agent as a smart triage nurse who has read every medical chart in the hospital. Instead of treating every patient the same way, they look at the current patient, remember how 10 similar patients were treated in the past, and make a smarter, more personalized decision. This saves time, reduces stress for patients, and ensures the expensive, invasive tests are only used when they are truly necessary.

Technical Summary

1. Problem Statement

Breast ultrasound is a critical frontline tool for cancer screening, particularly for women with dense breast tissue. However, current systems face significant challenges:

• Low Specificity: High rates of false positives lead to excessive diagnostic escalation.
• Unnecessary Biopsies: Approximately 60% of referrals result in benign outcomes, causing patient anxiety, psychological distress, and substantial economic/operational burdens on healthcare systems.
• Static Decision Logic: Traditional AI models rely on fixed thresholds and do not adapt their decision-making policies based on historical outcomes or accumulated clinical experience.
• Workflow Congestion: Excessive escalation of low-risk cases overwhelms radiologists, delaying care for genuinely suspicious patients.

The authors propose a solution that reduces unnecessary biopsies and diagnostic escalation while maintaining high sensitivity (detecting actual cancers).

2. Methodology: BUSD-Agent

The authors introduce BUSD-Agent, a novel experience-guided cascaded multi-agent framework. The system mimics clinical triage by splitting the workflow into two stages and utilizing a "memory bank" of past pathology-confirmed cases to adapt decision policies dynamically without retraining model parameters.

A. Cascaded Architecture

The framework consists of two distinct agents:

1. Screening Clinic Agent (Stage 1):

   • Role: Acts as a lightweight filter for all incoming ultrasound images.
   • Tools: Uses an ensemble of 14 pre-trained classification models (CNNs and Transformers like ConvNeXt, Swin, ViT, etc.).
   • Mechanism: An LLM orchestrator aggregates predictions via majority voting and confidence scores. It decides whether a case is "Benign/Normal" (Clear) or "Suspicious" (Escalate).
   • Goal: Filter out low-risk cases to prevent them from reaching the diagnostic stage.

2. Diagnostic Clinic Agent (Stage 2):

   • Role: Performs detailed analysis only on cases escalated by the Screening Agent.
   • Tools: Integrates a richer "Tool Zoo" including:
     • Multi-label predictors for radiological features (BIRADS, margins, shape, echogenicity, calcification).
     • A fine-tuned Vision-Language Model (VLM) for generating structured radiological reports.
     • Spatial tools: Object detection (bounding boxes), segmentation (lesion boundaries), and zoom/overlay utilities.
   • Mechanism: The LLM engages in a ReAct (Reason + Act) loop, iteratively reasoning over visual data, segmentation maps, and feature descriptions to make a final Biopsy Referral Decision.

B. Experience-Guided In-Context Adaptation

The core innovation is the Experience-Conditioned Policy Adaptation mechanism, which allows the agents to "learn" from past cases without parameter updates (fine-tuning).

• Biopsy-Grounded Memory Bank: A repository stores structured decision trajectories from past cases. Each trajectory includes:
  • Image embeddings.
  • Model predictions and confidence vectors.
  • Agent decisions (Escalate/Clear, Biopsy/No Biopsy).
  • Final pathology-confirmed outcomes (Ground Truth).
• Retrieval Mechanism: For a new query case, the system retrieves the top-$K$ most similar past cases from the memory bank using a joint similarity metric:
  • Screening Stage: Combines Image Embedding Similarity and Malignancy Confidence Vector Similarity.
  • Diagnostic Stage: Uses a richer feature space including radiological feature predictions (BIRADS, margins, etc.) and structured descriptors.
• In-Context Learning: The retrieved trajectories are formatted as "exemplars" and fed into the LLM's context window. This conditions the agent's current decision policy, allowing it to:
  • Identify which model outputs or confidence patterns were historically reliable for similar cases.
  • Adjust trust in specific tools dynamically.
  • Recalibrate escalation thresholds based on prior successes/failures.

3. Key Contributions

1. Cascaded Multi-Agent Framework: A two-stage system (Screening $\to$ Diagnostic) that mimics clinical triage, effectively separating low-risk cases from high-risk ones.
2. Experience-Conditioned Policy Adaptation: A novel method where agents adapt their behavior by retrieving and reasoning over past pathology-confirmed decision trajectories, enabling self-improvement without model retraining.
3. Comprehensive Tool Integration: Integration of diverse tools (segmentation, detection, radiological description generation) into an LLM orchestration layer for the diagnostic stage.
4. Extensive Benchmarking: The first large-scale benchmark of 10 proprietary and open-source Vision-Language Models (VLMs) in breast ultrasound screening, highlighting their instability compared to the proposed tool-assisted approach.

4. Experimental Results

The framework was evaluated across 10 diverse breast ultrasound datasets (e.g., BUSI, BUS-BRA, US3M, MICCAI 2022).

• Reduction in Escalation & Biopsies:
  • Diagnostic Escalation: Reduced from 84.95% (baseline) to 58.72%.
  • Biopsy Referrals: Reduced from 59.50% (baseline) to 37.08%.
• Performance Metrics:
  • Screening Specificity: Improved by 68.48% on average (significant reduction in false positives).
  • Diagnostic Specificity: Improved by 6.33%.
  • Sensitivity: Maintained or slightly improved (no substantial degradation in detecting malignancies).
  • Balanced Accuracy (BAcc): The proposed method achieved the highest BAcc (78.29% screening, 71.58% diagnostic) compared to state-of-the-art VLMs (GPT-4o, Gemini, Llama, etc.), which showed unstable trade-offs between sensitivity and specificity.
• Ablation Studies:
  • Retrieval Count ($K$): Performance peaked at $K=10$ retrieved exemplars.
  • Similarity Metric: The combined approach (Image + Confidence/Feature vectors) outperformed image-only or vector-only retrieval strategies.

5. Significance

• Clinical Impact: By significantly reducing false positives and unnecessary biopsies, the system addresses a major bottleneck in healthcare, reducing patient anxiety and operational costs.
• Adaptive AI: Demonstrates that LLM-based agents can dynamically adapt to new data distributions and refine decision policies using "experience" (retrieved trajectories) rather than static weights, offering a path toward safer, more reliable medical AI.
• Resource Efficiency: The cascaded approach ensures that computationally expensive diagnostic tools and specialist radiologist time are reserved only for cases with genuine suspicion, optimizing resource allocation in high-volume screening environments.