Overview

NOTE: This is an auto-generated report based on BACON aggregation structure (excerpt from https://arxiv.org/pdf/2505.14510).

This aggregation tree provides an interpretable decision logic for breast cancer diagnosis based on the WDSD dataset. The structure of the tree reflects a highly cautious approach, ensuring that both cell irregularity and size-related factors are strictly considered. The overall decision requires strong evidence from both aspects, following an aggregation path that demands multiple conditions to be met simultaneously for a diagnosis leaning towards malignancy.

Decision Logic Walkthrough

Top-level Aggregator: HHC — High Hyper-Conjunction (Extremely strict “must have all” behavior)

At the root, the model uses a High Hyper-Conjunction (HHC), which is among the strictest conjunctive operators. This operator requires that both child nodes provide strong signals simultaneously for the decision path to indicate a malignant cell. This means that even if one path shows high risk, the decision will not be affirmative unless the other path also confirms it.

First branch: Feature concave_points3

This feature measures the number of concave points (sharp indentations) on the cell nucleus border. A higher value suggests increased irregularity, which is a known indicator of malignancy.

Second branch: Aggregator HHD — High Hyper-Disjunction (Extremely tolerant “enough to have any of the most extreme cases”)

This operator is a highly permissive disjunction, meaning that the model is willing to accept any one of its child features showing extreme values as sufficient within this path.

Children of HHD:

  1. radius2 (Weight: 0.04) Reflects the radius of the nucleus. Lower weight suggests it is the least influential among this group.

  2. area1 (Weight: 0.11) Represents the area of the nucleus. Slightly more important but still relatively low influence.

  3. concavity1 (Weight: 0.34) Measures the degree of concave portions (dents) in the contour. Significant contributor in this group, highlighting that the degree of irregularity is critical.

  4. perimeter3 (Weight: 0.51) Captures the length of the nucleus border. Most influential feature in this group, suggesting that larger or irregular borders are strong risk indicators.

Clinical Interpretation

The aggregation tree reflects a clinically cautious and conservative diagnostic logic: The model requires both high concave points (concave_points3) and strong evidence from the size-related group (HHD) to consider the cell malignant concave_points3 plays a decisive role, acting as a gatekeeper that must be high regardless of what happens in the size group. The HHD branch, although permissive in combining its inputs, is treated strictly at the root level due to the HHC operator. This means that the model tolerates multiple possible manifestations of size-related abnormalities but does not allow them to override the need for high irregularity. perimeter3 and concavity1 are the standout contributors within the HHD group, confirming that larger perimeter and irregular borders are strong secondary indicators.

Key clinical takeaways:

Irregular cell shape (concave points and concavity) is mandatory for malignancy prediction