The tracebloc Playbook: How to Enhancing Breast Cancer Screening with Secure AI Model Evaluation

Tracebloc is a tool for benchmarking AI models on private data. This Playbook breaks down how a team used tracebloc to benchmark AI models on their breast cancer dataset and discovered which model truly delivered the best results. Find out more on our website or schedule a call with the founder directly.

Why Model Performance Matters

Every missed tumor risks lives and costs money and every false alarm triggers follow-up tests. Using tracebloc, a academic research institute uncovered which object detection model truly performs under pressure, saving over €500k a year compared to the next best approach.

Use Case Description

Dr. Elena Fischer, Lead Data Scientist at a major university hospital in Berlin, is tasked with enhancing diagnostic quality in breast cancer screening: Missed tumours lead to delayed treatment and higher mortality, while false positives drive unnecessary biopsies and costs.
Radiologists at the hospital currently achieve 96% sensitivity with ~5% clinically acceptable false positives, but manual review is time-consuming and subject to fatigue. The goal is to provide an AI assistant to support radiologists, especially in the pre-diagnostic phase—helping to prioritize scans and reduce false negatives.

Key requirements:

The model must run on-premises or close to the edge
It must reach ≥97.5% sensitivity in real-world testing at <5% fixed false positive rate (FPR)
It must integrate with PACS and respect all GDPR and MDR regulations

The hospital has access to a valuable internal dataset: 50.000 annotated X-ray scans (malignant/benign). While developing an in-house model remains a potential path, estimated to take 12 to 18 months, Elena decides to first evaluate the current state of the market. By conducting a structured assessment of 3 CE-marked AI vendors, she aims to understand the current state of the art, explore available commercial solutions, and gain a clearer picture of the associated costs, performance, and integration effort. This approach also provides the opportunity to access pretrained models that may generalize better and exhibit less institutional bias than a model trained solely on internal data.

Step 1: Vendor Model Metrics

Each vendor submits commercial and technical proposals:

VENDOR	CLAIMED SENSITIVITY / SPECIFICITY	COST PER IMAGE	INFRASTRUCTURE LOAD
A	95% / 94%	€0,50	Low
B	96% / 95%	€1,00	Moderate
C	97% / 96%	€1,80	High

Elena and her team prioritize sensitivity due to the high risk of missing early tumours, but they cap the false positive rate at 5% to avoid excessive misdiagnoses.

Step 2: Secure Evaluation and Fine-Tuning

Using tracebloc, Elena sets up secure sandboxes within the hospital’s IT infrastructure. Vendors do not get access to the raw data. Instead, they are invited to fine-tune their models securely on-prem, ensuring full data protection and regulatory compliance.

Each vendor is given access to:

40.000 images for training
10.000 held-out images for testing
Predefined benchmark metrics: AUC, sensitivity, specificity, and inference time

After baseline evaluation, vendors fine-tune their models and submit new versions for testing. The results reveal a dramatic difference between claims and real-world performance on the hospitals proprietary data.

Observed Results After Testing

VENDOR	CLAIMED SENSITIVITY	BASELINE SENSITVITY @ 5% FPR	SENSITVITY @ 5% FPR AFTER FINE-TUNING
A	0,95	0,87	0,89
B	0,96	0,91	0,691
C	0,97 ✅	0,92 ✅	0,978 ✅

Surprise outcome: Vendor A did not match its claimed sensitivity. Vendor B held its promise but was outperformed by Vendor C who significantly improved after secure fine-tuning.

Step 3: Business Case

Assumptions:

Annual volume: 50.000 X-ray scans
Cancer prevalence: 0.5%
Human sensitivity baseline: 4%, i.e. 10 missed cancers and ~2500 falsely diagnosed cases
Cost per missed cancer: €100.000 (delayed treatment)
Cost per false positives: €2.000 (biopsy, imaging, stress)
AI usage: Full batch inference with radiologist-in-loop

APPROCH	ESTIMATED SENSITIVITY @ 5% FPR	MISSED CANCERS	COST DUE TI MISSED CANCERS	AI COST	TOTAL COST
Human Only	~0,96	10	€1.000.000	€0	€1.000.000
Vendor A	0,890	28	€2.800.000	€25.000	€2.825.000
Vendor B	0,961	10	€1.000.000	€90.000	€1.090.000
Vendor C	0,978	~6	€600.000	€50.000	€650.000

False positive costs (~€5M/year) are held constant across all approaches (5% FPR), so focus can be set solely on the impact of missed cancers.

Step 4: Vendor Selection and Strategy

After secure on-prem fine-tuning, Vendor C’s model achieved the highest sensitivity (0.978). The best-performing setup emerged from a hybrid strategy, where the AI model pre-sorts scans and flags high-risk cases for radiologists to review. This setup reduces diagnostic errors by over 60%,improves efficiency, and maintains full clinical oversight—radiologists stay in control at every step.

The hospital selects the Human + AI configuration, which delivers both strong medical outcomes and a compelling business case:

STRATEGY	ESTIMATED SENSITIVITY	MISSED CANCERS	COST DUE TO MISSED CANCERS	TOTAL ANNUAL COST
Human Only	~0.960	10	€1,000,000	€1,000,000
Human + Vendor C	~0.985	~4	€400,000	€450,000 ✅

Estimated annual savings: €550.000

The hospital proceeds with a 3-month pilot, including:

On-prem deployment and PACS integration
Radiologist-in-the-loop configuration
Monthly performance validation and audit reporting

Disclaimer:
The persona, figures, performance metrics, and cost calculations in this case study are illustrative and based on fictionalized inputs designed to mimic real-world scenarios. They are intentionally kept at a high level to make the concepts easier to understand and communicate. These do not represent actual clinical results, vendor performance, or contractual terms, and are intended solely for strategic discussion and conceptual exploration.

AI-Powered Breast Cancer Screening and Image Classification for Radiology