The Story in the Data

This project started with a question: how is breast cancer actually diagnosed? I was in a machine learning course at the University of Nottingham, working with a breast cancer dataset. At first, I focused on accuracy of the model, adjusting models, testing features, trying to get better results.

But the deeper I got, the more I wanted to understand what the data couldn’t tell me; the reasoning behind clinical decisions.

When someone close to me was diagnosed with breast cancer, I saw how much uncertainty surrounds the diagnostic process. It made me curious about how design and machine learning could bring more clarity and confidence to real-world diagnosis. That experience stayed with me, and I got the chance to pick the project back up with the University of Washington to explore those possibilities further.

"Misdiagnosis and overdiagnosis remain key challenges in breast cancer imaging, where conventional mammography may fail to detect lesions."

^{Thomassin-Naggara et al. (2024)}

Why we did this?

One of the biggest challenges we encountered was rooted in the complexity of diagnosis. Despite the availability of tumor metrics and imaging data, determining whether a tumor is benign or malignant isn’t always clear-cut. Reading imaging data like mammograms or ultrasound scans is highly nuanced and often subjective. Small differences in shape, margin, or density can be hard to interpret and may lead to inconsistent assessments across clinicians, especially in borderline or low-quality cases.

Secondary research

We started by looking into existing research to understand why diagnosing breast cancer is often so complex. Radiologists interpret features like shape, margin, and density differently, and even small changes can lead to different outcomes. This is especially true in borderline cases or when images aren’t clear. These insights helped us focus on where and why the problem exists and to design tools that support clinical judgment.

Human error factor

• False negatives in mammography range from 12–30%, depending on case complexity and image quality
• Up to 50% of cancers in dense breasts may go undetected without additional imaging
• Benign findings can be flagged as dangerous, leading to unnecessary biopsies and anxiety

Gaps in existing tools

• AI models are often hidden behind a blackbox where most don’t explain why a case is flagged, limiting clinical trust‍
• Generic outputs like “malignant: 84%” may lack clinical value and doesn't answer "why?"‍‍
• Tools often miss real-world workflows – a gap in human-centered design

Opportunities for innovation

• Tools to show "why" a tumor is flagged by surfacing the specific features influencing the decision
• Radiologists want tools that support clinical judgment, not to automate the process
• Outputs should adapt to case-specific contexts like dense breast tissue or borderline features

Machine Learning Model

The machine learning model was built early on, during my time at University of Nottingham, as a way to explore how tumor characteristics could predict malignancy and patterns, especially whether those patterns aligned with how radiologists make decisions. At that stage, I didn’t know this would evolve into a design project. But training the model helped uncover which features were most influential, which later became critical input for designing an interface that could surface meaningful, case-specific insights and support clinical reasoning.

quantitative data exploration

After deciding to turn this into a design project, I revisited the model through deeper quantitative analysis to unpack how its predictions worked in detail. I wanted to explore which features to emphasize, how uncertainty showed up in the data, and where edge cases might cause confusion. These visualizations helped shape the design direction, especially around what to prioritize, how to handle ambiguity, and how to build trust through clarity and transparency.

Triangulated Research-to-design Matrix

Before designing the interface, I needed to understand how diagnostic decisions break down – in models, in data, and in clinical workflows. I triangulated three methods to create a research-to-design matrix that helped validate patterns across sources and identify high-confidence insights. The matrix surfaced ten core findings that revealed where errors happen, what users actually need, and how AI predictions can be made more interpretable. These insights became the foundation for every UI decision that followed.

By combining model behavior, pattern analysis, and literature on breast cancer diagnostic processes, I mapped out the most impactful pain points: unreliable feature weighting, lack of transparency, edge cases, and cognitive overload. Each design decision below directly addresses these breakdowns with targeted interface responses.

Enhanced Diagnostic Precision with Mean & Worst Metrics

Radiologists don’t just look at a tumor’s ‘average’ size, they also zero in on its single most abnormal spot in a single patient. Missing that one extreme region can lead to under-diagnosis.

"Existing breast imaging studies reported the entropy, mean, minimum, and maximum as important features."

^{Lee et al. (2020)}

MEan

• Calculated by sampling X (e.g. radius, area, texture) at dozens or hundreds of points on the same tumor and taking their average. Reflects the lesion’s typical size, shape, or heterogeneity.
• Without mean, radiologists lose important context – every lesion has a baseline appearance and the model detects whether this appearance is benign or malignant.

Worst

• From the very same measurements, radiologists select the largest (or near-largest) value. Highlights the single most abnormal “hot spot” that may warrant targeted biopsy.
• Worst metric measurements prevent dangerous outliers from hiding in the average, ensuring that even small but aggressive regions of the tumor are flagged for further clinical attention.

I co-built the interface using Streamlit and iterated directly in code (using vibe-coding), guided by user needs and model behavior. Streamlit allowed me to maintain full control over the model logic while rapidly prototyping interfaces that stayed true to the algorithm’s outputs. Unlike visual design tools, Streamlit let me directly connect model predictions with interface elements, making it easier to test ideas in real time, adjust how probabilities were framed, surface uncertainty, and experiment with interactive features like sliders, graphs, and confidence estimates.

V1 - Basic Inputs, No Guidance

• The first version was a straightforward input form where users manually entered four tumor metrics to generate a prediction. While functional, it offered no interpretive support, making the experience feel opaque and limiting users’ ability to trust or make sense of the output.

V2 - Sliders and Contextual Info

• The second version introduced interactive sliders, population averages, and brief metric descriptions to improve usability and reduce friction. This helped users understand what they were adjusting, but the model’s reasoning was still unclear and users couldn’t easily connect inputs to outcomes.

V3 - Transparent and Decision-Supportive

• The final version focused on interpretability and trust. It added confidence labels, similar-case comparisons, and a feature-level visualization showing how each metric influenced the result. These changes transformed the tool into a decision-support interface that aligned more closely with user needs (radiologists) and mental models of breast cancer tumor diagnostics.

Reflections: Merging Code, Data, and Design

Working on this project was full of surprises. At first, weaving HTML and CSS into Python felt unfamiliar, but that challenge helped me see how code underpins every interface decision. Diving into quantitative research and creating data visualizations revealed diagnostic patterns that static mockups couldn’t capture. Turning my own ML model into an interactive breast cancer app emphasized the power of combining data and design. Watching complex tumor metrics become clear, actionable visuals for radiologists was a true highlight, it was proof that solid research and thoughtful coding can drive meaningful user experiences.

Roadmap: From Prototype to Practice

I’m eager to move this prototype into real-world workflows by recruiting radiologists, medical students, and interdisciplinary care teams for hands-on usability testing. Watching how diagnostic teams navigate the interface will uncover pain points and guide refinements to how we surface feature importance and communicate uncertainty. I’ll iterate on the design based on their feedback, then pilot the app in both clinical and educational settings to measure its impact on diagnostic confidence, learning outcomes, and overall workflow efficiency. These steps will ensure the tool truly integrates data-driven insights into everyday diagnostic practice.