A preliminary study proposed entropy-weighted hybrid pooling for CNN classification of 9,016 public breast-ultrasound images. The three-block model averaged 93.98% accuracy and 0.987 AUC, slightly above max pooling, but hybrid pooling did not win in the four-block model, so general superiority is not established.
Key findings
- In the three-block setting, hybrid pooling reached 93.98±1.72% accuracy and 0.987 AUC versus 92.72±0.85% and 0.9815 for max pooling. In a reported single four-block run, hybrid reached 92.90%, below max pooling at 94.79%, indicating architecture-dependent rather than consistent benefit.
Why this matters globally
Data-adaptive pooling may help compact medical-image models combine feature scales, but clinical value requires external datasets and safety-oriented evaluation.
Thai researcher contribution
The Suranaree team proposed an architectural modification and explored when it helped, contributing a Thai AI-methods study relevant to medical imaging.
Limitations to consider
Only one public dataset was used, and patient-level versus image-level splitting is unclear, creating leakage risk. There was no external validation, calibration, radiologist comparison or workflow assessment, and some results came from a single run.