Quantitative analysis of biofilm images can be framed as dense prediction under severe class imbalance and scarce supervision. At whole-slide resolution, only about 2% of image regions carry annotations. Microsegmentation therefore requires estimating a pixel-wise labeling function by risk minimization from sparse, noisy labels, and the bottleneck is the quality and quantity of labeled data rather than model capacity. Progress in machine learning has repeatedly been driven by shared, well-annotated data. In speech recognition, the common evaluation paradigm, with shared corpora, fixed metrics, and successively harder tasks, steadily reduced error rates for decades. Many scientific disciplines lack comparable infrastructure. However, vibe coding tools such as Claude.ai are changing the way we develop data and the associated technology.
In this talk, we discuss several architectures we have developed that were fine-tuned from pretrained weights: residual CNNs (ResNet-18), compound-scaled CNNs (EfficientNet B0/B7), and attention-based models (ViT-16/32). Performance is measured by mean Dice (MDICE), the overlap between predicted and annotated regions with background excluded. Results are compared against model complexity (5.3M to 88.2M parameters) and training and decoding cost. The smallest model, EfficientNet-B0, outperforms the larger transformers on both datasets, reaching MDICE of 51.10 and 46.02 on the evaluation sets. This suggests that inductive bias and parameter efficiency matter more than raw scale when labeled data are limited.
We also show that segmentation models trained on partially annotated data approach the accuracy of those trained on fully annotated data. Partial annotations are converted into fully annotated training samples by assembling collages of subimages, which reduces annotation cost and effort. Finally, we discuss how evaluation granularity limits what can be demonstrated: small dev and eval sets cannot resolve differences between models at a strict operating point. We argue that standardized splits, pre-registered metrics, and confidence intervals are prerequisites for reproducible comparison, and we describe the infrastructure to support them. One such recent development is the establishment of the Biofilm Imaging Library (http://biofilmimagelibrary.org/).