Three automated devices which utilised neural networks for smear analysis have been subjected to extensive multicentre trials and subsequently approved by the American Food and Drug Agency for screening cervical smears. These were:
The Papnet and AutoPap systems were developed in the 1990s have recently been replaced by the Tripath Focalpoint system. All three are interactive systems ie. they selected smears for manual review by the screener. They were approved initially for quality control or supplementary screening of cervical smears. Susequently they were approved for primary screening.
Papnet was introduced as a pre-screening method of conventional Pap smears. The slides were read through the Papnet scanner which selected images of interest. These were then reviewed on a computer screen by a scientist. Papnet used neural net computer technology to analyse the complex nature of the conventional Pap smear.
AutoPap was designed to also pre-screen conventional slides. Conventional computer technology was used. Images were not presented for review but rather slides with the most severe abnormalities were grouped together and these slides were then completely screened by scientists. The percentage to be completely screened could be altered but whichever component was not re-screened was archived and was not ever reviewed by a scientist or pathologist.
The AutoPap system received FDA approval in May 1998 for use in primary screening of Pap smears for cervical cancer. The Autopap system scanned slides of conventionally prepared Pap smears and ranked specimens according to their degree of abnormality. The system was designed to look for abnormalities slide by slide and to rule out the 25% of slides with the lowest risk. These slides were automatically excluded from the list of those requiring manual microscopic review thus reducing the screener’s workload by 25%.
Papnet was a commercial neural network-based computer program for assisted screening of Pap (cervical) smears.
The smears were analysed using a combination of algorithmic and neural network programs and 128 images of the most abnoraml looking cells or cell groups were selected for inspection by the screener.
The images were stored on compact disc and viewed by the screener on a video monitor in the laboratory. The screener triages the images and decides whether the slide is negative or requires manual review. Those slides which were triaged negative were not subjected to manual microscopic review.
Both Papnet and Autopap were tested in extensive multicentre trials which compared automated screening and manual screening of the same slides.
The trials found that the automated systems were at least as sensitive as manual screening and that more smears could be analysed per unit of time.
Due to high development costs the systems were not found to be cost effective for use by cytology laboratories processing less than 50,000 smears per annum which excluded all but a few laboratories in the USA and Europe. Consequently they were not commercially viable.
Both the Papnet and the AutoPap technology were purchased in some part by a third group, SurePath-Autocyte and are no longer commercially available. However, in their time, they represented a breakthrough in automated cervical screening.
TriPath FocalPoint™; This imaging system is an automated cervical cytology screening device which is currently available . It is intended for use in initial screening of cervical cytology slides. The FocalPoint identifies up to 25% of successfully processed slides as requiring no further review. The FocalPoint also identifies at least 15% of all successfully processed slides for a second manual review.The device is intended to be used on both conventionally-prepared and SurePath™ (formerly AutoCyte®PREP) cervical cytology slides. For both preparation methods, the device is intended to detect slides with evidence of squamous carcinoma and adenocarcinoma and their usual precursor conditions.
Computer Based Image Analysis
Vision is a complex process of the eye receiving images as light and then the brain interpreting the images in the context of signals from other senses as well as the conscious and subconscious parts of brain activity. In cytology, we train our interpretive ability by looking at hundreds or thousands of samples, reading, being taught etc. We make mistakes and learn from them. Such processes are difficult to simulate using computers but the development of LBC and the production slides with cells as monolayers has given us the possibility for semi-automation.
We can take an example of the one of the criteria we use for diagnosis in cytology – hyperchromasia. When we look at normal squamous epithelial cells we know what the nuclear appearance should be. The nuclear membrane is thin and the chromatin shows a pattern of finely divided heterochromatin and euchromatin. If we see areas within the nucleus where the heterochromatin appears to be uneven and is concentrated in some parts of the nucleus and not others, we describe that as chromatin clumping and interpret it as a sign of abnormality.
Digital cameras “see” images by measuring the light intensity and colour properties being received by their electronic sensor elements. We refer to these as pixels (pixel = PIcture ELement) and the values from light/dark and colour can be measured and stored by a computer attached to the camera. If we place our stained cytology samples in an apparatus which has lenses and a digital light sensor (camera) we can “train” the computer to react to chromatin clumping as well as some of the other criteria we use, such as nuclear size, form etc. For example, heterochromatin can be detected as groups of dark pixels, which are different when compared to the finely distributed chromatin present in normal nuclei.
These “rules” for recognition of abnormal structures are called algorithms and are incorporated in the systems marketed by the companies mentioned above.
Surepath uses a system called FocalPoint®. This analyses the samples using a series of algorithms as described above and assigns a score to the sample. The sample is then graded into a group called No Further Review (NFR) or into one of 5 risk categories. The purpose of this is to make it unnecessary for a trained person to look at the NFR category. They can instead concentrate on looking at the slides graded as some level of risk of abnormality. The operator is guided to the areas containing the cells of interest (Fields of View/FOV) which have been detected by the system.
Cytyc markets the ThinPrep® Imaging System which functions in a similar way to the SP system. It has 22 FOV – which the system has ranked as being the most significant and that an experienced person should look at.
Evaluation of automated systems for the analysis of cervical smears DV Coleman 1998 Cytopathology 9,359-368
Automated cervical cancer screening edited by Heinz Grohs and OAN Hussain (1994) Igaku-shoin New York
Cervical screening programmes: can automation help? Evidence from systematic reviews, an economic analysis and a simulation modelling exercise applied to the UK. Willis BH, Barton P,Pearmain P,Bryan S,Hyde C. Health Technol Assess. 2005 Mar;9(13):1-207, iii.