Urinary tract infections (UTI) are a common disorder. Its diagnosis can be made by microscopic examination of voided urine for cellular markers of infection. We present a dataset containing 300 images and 3,562 manually annotated urinary cells labelled into seven classes of clinically significant urinary content. It is an enriched dataset with samples acquired from the unstained and untreated urine of patients with symptomatic UTI. The aim of the dataset is to facilitate UTI diagnosis in nearly all clinical settings by using a simple imaging system which leverages advanced machine learning techniques. 

Data acquisition 

300 urine samples were obtained from patients with symptomatic UTI between April and August 2022 from a specialist LUTS outpatient clinic in central London. Urine samples were collected as natural voids and processed on-site within one hour to mitigate cellular degradation. Brightfield microscopic examination (Olympus BX41F microscope frame, U-5RE quintuple nosepiece, U-LS30 LED illuminator, U-AC Abbe condenser) was performed at x20 objective (Olympus PLCN20x Plan C N Achromat 20x/0.4). A disposable haemocytometer (C Chip™) was used for enumeration of red cells (RBC), white cells (WBC), epithelial cells (EPC), and the presence of other cellular content per 1 µl of urine by two experienced microscopists.

Images were acquired using the aforementioned brightfield microscope using a 0.5X C-mount adapter connected to a digital colour camera (Infinity 3S-1UR, Teledyne Lumenera). Images were taken in 16-bit colour in 1392 x 1040 .tif format using Capture and Analyse software. An enriched dataset approach was taken to maximise urinary cellular content in the acquired images. Such data curation was also necessary to overcome class imbalance. Daily Kohler illumination and global white balance was performed to ensure consistency in image acquisition. 

Dataset annotation

300 images were acquired and manually annotated by first identifying cells of interest as a binary semantic segmentation task. Individual pixels were dichotomously labelled as either informative cells, foreground, or non-informative background. Non-informative background was further constrained by including unidentifiable cells, such as debris or grossly out-of-focus particles. Binary annotation was initially performed using ilastik, an open-source software using a Random Forest classifier for pixel classification, then manually refined at the pixel level to ensure accurate semantic segmentation. This produced a binary mask in 1392 x 1040 .tif format for each corresponding raw colour image. 

Objects of interest were then manually labelled by two expert microscopists into one of seven clinically significant multi-class categories: rods, RBC/WBC, yeast, miscellaneous, single EPC, small EPC sheet, and large EPC sheet. This produced a multi-class mask in 1392 x 1040 .tif format with a label as pixel value from 0-7, where 0 is background (Table 1). 

Data structure 

The dataset is organised into three root folders: img (image), bin_mask (binary mask), and mult_mask (multi-class mask). Each folder has 300 files in .tif format and labelled with an incremental number.

Table1

Folder Files  Objects  Count Pixel Values

img 300 Raw data 0-255 bin_mask  300 Background/Foreground 0/1 mult_mask  300 Background/Class 0 Rod 1697 1 RBC/WBC 1056 2 Yeast 41 3 Miscellaneous  550 4 Single EPC 182 5 Small EPC sheet 26 6 Large EPC sheet  10 7

                        Total                  3562