This paper presents an algorithm that can be used to assign laterality – Left or Right and image type – Mediolateral oblique (MLO) or Craniocaudal (CC) for newly digitized mammographic films. This information is vital for correct image display on reporting workstations. The order and the position of the image on the diagnostic display is controlled by hanging protocols, which are based on the Laterality/View type information supplied in DICOM header for each image being displayed. Images that originate from a digital modality usually have this information supplied by the modality, while digitized analogue films do not, as the image digitizer has no means to analyze the film content. To overcome this issue, the operator who digitizes images has to assign missing information manually. This approach is error prone and time consuming.

1. Statistical Image Recognition Techniques. A digitized imaged is compared to a set of “typical” images from an image library. During each comparison a correlation factor is calculated, based on similarity between the image pixels of the analyzed image and pixels from the library image. The Laterality/View type information from the image with higher correlation factor is assigned to the analyzed image. This most flexible approach can work well for any body part or for a mixture of different images. The drawback is unsatisfactory performance, as each image is compared against a large number of targets, and each comparison requires substantial processing resources.

2. Image Vectorization. Body part shapes from the image being analyzed are approximated with a set of curves. The decision about Laterality/View type is based on comparing spatial characteristics of the curves. This approach is flexible as well, however it requires complex mathematical analysis of the image pixel data and could be quite slow.

3. Image Pixel Analysis. This approach is based on the calculation of various image metrics, based on the image pixel data. The algorithms from this group are generally very fast, however image recognition accuracy is not as high as in group 1 or 2. These algorithms work very well when the number of possible image groups is limited to a small number of choices.

The algorithm presented in this abstract falls in the third category. The bulk of the images taken within the population screening environment falls into just four groups: Left Mediolateral oblique (L-MLO), Right Mediolateral oblique (R-MLO), Left Craniocaudal (L-CC), and Right Craniocaudal (R-CC). At the same time the images are digitized in large batches by non-clinical personnel: productivity is important, and the possibility of human error is high. It makes it the ideal environment to use an algorithm from the third group.

Image conditioning:

1. Normalize brightness/contrast as the images acquired from different sources have a large range of average optical density and contrast.

2. Increase image contrast and brightness by a certain value to produce an almost monochrome image.

3. Convert the image to 8-bit for processing speed. Mammographic images are usually 12 or 16 bit. Image quality has no bearing on recognition accuracy, converting the image to lower resolution decreases image size and greatly increase processing speed.

4. Remove dust and scratches by increasing a size of black pixels.

5. Crop the top and bottom of the image to reduce impact of stray light recorded on a film.

Image Laterality:

1. Divide the image into two segments of equal width along the vertical image axis. See Fig. 1.

2. Sum the number of pixels having value above a certain gray threshold in each segment identified in the previous step.

3. If number of pixels calculated in step 2 from the right segment exceeds the number of pixels from the left segment, the image is assigned Right Laterality. Otherwise, the image is assigned Left Laterality.

Fig.1 Determining Image Laterality.

Image View Type:

View type identification algorithm for the left image mirrors the algorithm for right one. Following is a sequence of steps for the left image.

1. Detect skin contour.

2. Identify vertical coordinate of a rightmost point on a skin contour. As the skin contour for digitized images may not be well defined, calculate this value by averaging vertical coordinate over a number of pixels located at equal distance near the skin edge.

3. Sum the number of the image pixels having a value above a certain gray threshold from the region above and below the vertical coordinate determined in previous step.

4. Calculate asymmetry index by comparing values calculated in step 3. Ia= (Na- Nb)/ Na. See Fig 2.

5. Make a decision – MLO, CC or indeterminate. Ia thresholds for MLO and CC could be determined by applying the algorithm over a series of images of well known type. In our experiments images with the Ia value of less than 0.2 were of CC type while images with Ia value greater than 0.4 were MLO. Ia values between 0.2 and 0.4 indicate that the image type could not be reliably detected.

6. If the outcome from the previous step is indeterminate, test for the presence of pectoral muscle. The most common reason for an indeterminate outcome is that the rightmost point on a skin contour may not be near the nipple area for images of small breast and comparatively large pectoral muscle. See Fig 3.

7. Select a region of approximately quarter of the image width at the top of the image and calculate number of pixels having value above certain gray threshold.

8. Select a region of approximately quarter of the image width at the centre (half height) of the image.

9. Calculate second asymmetry index by comparing values calculated in steps 7 and 8. Iaa= (Nc- Nd)/ Nd

10. Iaa below certain value (3.0 in our case) indicates presence of an pectoral muscle which is specific to MLO images only.