The problem — Automated ingestion pipelines process thousands of tiles per run. Manual quality review does not scale, and classical no-reference metrics (NIQE, BRISQUE) are trained on natural-scene statistics that transfer poorly to fluorescence and bright-field microscopy. A pipeline that relies on them will pass blurry fields, saturated channels, and artifact-laden tiles silently into segmentation and profiling, degrading every downstream measurement without a warning signal.
What it is / how it works — μDeepIQA is a deep CNN trained on microscopy-specific quality ratings rather than on natural-scene benchmarks. Its key architectural choice is global-to-local prediction: the model produces both a whole-image quality score and per-patch scores, so it can identify which sub-regions are degraded — out-of-focus corners, saturation halos, stripe artifacts — rather than merely flagging whole images. This patch-level output maps directly onto tile-based workflows: a tile that is spatially heterogeneous can be partially accepted or re-queued rather than discarded wholesale. The model is designed for fast inference and for robustness to acquisitions that fall outside the training distribution, a critical property when instruments, objectives, and fluorophore choices vary across client sites.
μDeepIQA is the companion to the Multi-Marker Similarity (MMS) approach from the same authors, which decomposes quality into interpretable physical markers (resolution, SNR, contrast) for reduced-reference assessment when a coarse reference is available. Together they cover the two practical regimes of the image-space quality level: reference-free screening (μDeepIQA) and interpretable reduced-reference auditing (MMS).
Where it breaks — No-reference predictions are only as reliable as the training distribution. Performance degrades on acquisition modes that differ substantially from the training corpus — novel objectives, unusual fluorophores, 3D z-stack cross-sections, or label-free phase-contrast modalities. In those regimes the model's confidence may be high while its predictions are miscalibrated, which is more dangerous than an obvious failure. Additionally, no-reference quality scores measure image-space fidelity; they say nothing about whether the acquisition is task-adequate — a moderately soft field may still yield reliable morphological features, while a visually sharp field may contain a staining artefact that poisons the readout. μDeepIQA belongs at the image-space level of Image Quality Is Not One Number; it is not a substitute for run-level QC, readout-level validation (percent-replicating, mAP), or the operational verdict of a QC-aware report.