Investigators at the Hauptman-Woodward Institute (HWI) and the Ontario Cancer Institute are collaborating in the development of automated image analysis methods for detecting the presence of diffraction-quality crystals produced by the microbatch-under-oil experiments conducted in HWI's High-Throughput Screening lab. The first requirements for developing such methods are finding good ways to classify the outcomes of crystallization experiments and to improve the image acquisition and storage steps.

Image acquisition

HWI lab technicians must currently guide the imaging robot to register the corner wells of a plate before imaging. To automate this step, we have adapted our well-registration software to identify whether a well is in view of the camera and, if so, to give precise coordinates of the well centre (even in the case where the well is over 75% out-of-frame). This software, to be incorporated into the next version of HWI’s camera-control software, will save ~10 man-hours per week in setup time.

Compressed image storage

Precise record of the contents of each crystallization trial is required both for archiving and for image analysis. HWI currently archives crystallization images as RAR-compressed batches of TIFF files, requiring nearly two GB of storage for a single plate imaged at seven time points. To provide efficient, crystallization-specific, lossless image compression, we have investigated the use of JPEG 2000, a format that allows for lossless coding of arbitrary regions of an image. Using our well-registration algorithm to identify a 320-pixel diameter circle covering the entire well, we tailored JPEG-2000 compression to compress the well region without lossless while compressing the exterior without distortion constraints. The compressed output is fully JPEG-2000 compliant, and it is mathematically guaranteed to be the smallest stream that a JPEG-2000 compressor can achieve without altering the interior of the well. In a study of 279,552 images (182 plates), the JPEG-2000-compressed archive was 25.26% the size of the RAR-compressed TIFFs — smaller than a lossless compressed crop of the 320x320 square region of interest.

Image classification

We aim to classify automatically all images generated by the HWI robotic imaging system and to eliminate the need for a crystallographer to search hundreds of images for crystal hits. Based on a library of truth data of 276,480 images (5,553 with crystals), we instructed eight individuals to select the most relevant image features from a set of 840 computed by our image analysis software, modeling crystal-positive/crystal-negative images as probability distributions in a multidimensional feature space. Using an ensemble of such models, we can tailor output for high recall (68%) or high precision (41%) of crystal outcomes, with a mean 95% overall accuracy of image classification. Typically, a crystal hit (if present) is found in the first three top-scoring images on a plate. 74% of all plates have a crystal in the top 10; 95% have a crystal in the top 100 images.

The next step is to focus on developing a classification system with the following seven categories:

• Clear. A drop may have surface blemishes and still be considered clear.

• Phase separation. Presence of two distinct liquid phases with an appearance of oil droplets in vinegar.

• Precipitate. Light to heavy, uniform to structured

• Skin. A wavy blanket over the drop.

• Crystal. An object with well-defined edges. A starting point for an optimization experiment.

• Garbage. Images with diverse defects, out of focus, drops failed to merge, etc.

• Unsure. When it isn't obvious how to place an image in any of the other six categories.