Effectiveness of automated cervical cytology rescreening using the AutoPap 300 QC System

Performance characteristics of an artificial intelligence based on convolutional neural network for screening conventional Papanicolaou-stained cervical smears

Background

The conventional Papanicolaou-stained cervical smear is the most common screening test for cervical cancer. The sensitivity of the test in detecting abnormal cells is 67-75% in various studies. Owing to the volume of smears at cancer screening centres, significant man-hours are expended in the test. We have developed a software program for identification of foci of abnormal cells from conventional smears. We have chosen the convolutional neural network (CNN) model for its efficacy in image classification.

Methods

A total of 1838 microphotographs from cervical smears, containing 1301 'normal' foci and 537 'abnormal' foci were included in the study. The data set was split into training, testing and validation sets. A CNN was developed in the Python programming language. The CNN was trained with the training and testing set. At the end of training, 94.64% accuracy was achieved in the testing set. The CNN was then run on the validation set (441 images).

Results

The CNN showed 94.28% sensitivity, 96.01% specificity, 91.66% positive predictive value and 97.30% negative predictive value. The CNN could recognise normal squamous cells, overlapping cells, neutrophils and debris and classify the focus appropriately. False positives were reported when the CNN failed to recognise overlapping cells (2.7% microphotographs). It could correctly label cell clusters with high nuclear cytoplasmic ratio and hyperchromasia. In 1.8% of microphotographs, a false negative was reported.

Conclusion

The CNN showed 95.46% diagnostic accuracy, suggesting potential use in screening.

Molecular Basis for Advances in Cervical Screening

Human papillomaviruses (HPVs) cause cervical lesions, which can, in some instances, progress to high-grade neoplasia and cancer. Around half a million cases of cervical cancer occur each year, with most occurring in developing countries where cervical cancer is a major cause of cancer-related death. The reduction in cervical cancer incidence in developed countries is largely attributed to the introduction of cervical screening. Cervical screening currently depends on the identification by cytology of abnormalities in cells taken from the surface of the cervix. The standard Pap test was developed >50 years ago, and despite modifications, still forms the basis of the test currently in use in most routine screening laboratories. Advances in our understanding of the molecular mechanisms that lead to the development of cervical cancer have been slow to impact on screening, despite the relatively high false-negative rates that can be associated with the conventional Pap smear. Improvements in screening strategies fall into a number of categories. Methods that improve cell presentation and attempt to eliminate artefacts/obscuring debris can be combined with image analysis systems in order to enhance diagnostic accuracy. Such approaches still rely on cytological evaluation and do not incorporate advances in our knowledge of how HPV causes cancer. By contrast, markers of virus infection or cell cycle entry, particularly those that offer some degree of prognostic significance, may be able to highlight abnormal cells more reliably than cytology, and could be combined with cytology to improve the detection rate. Our understanding of the molecular biology of HPV infection and the organization of the HPV life-cycle during cancer progression provides a rational basis for marker selection. The general assumption that persistent active infection by high-risk HPV types is the true precursor of cervical cancer provides the rationale for HPV DNA testing in conjunction with enhanced cytology, while the development of RNA-based approaches should allow active infections to be distinguished from those that are latent. The detection in superficial cells of marker combinations at the level of RNA or protein has the potential to predict disease status more precisely than the detection of markers in isolation. There is also a need for better prognostic markers if the predictive value of screening is to be improved. The potential to control infection by vaccination should reduce the incidence of HPV-associated neoplasia in the population, and this may cause a change in the way that screening is carried out. Nevertheless, the lack of a therapeutic vaccine, and the difficulties associated with eliminating infection by multiple high-risk HPV types, means that some form of screening will still be required as a preventive measure for the control of cervical cancer for the foreseeable future.

A novel automated screening and interpretation process for cervical cytology using the internet transmission of low-resolution images

BACKGROUND

Transmission over the Internet of low-resolution images acquired by automated screening of cervical cytology specimens has the potential to provide remote interpretation and, hence, centralization of a cytology workforce.

METHODS

Liquid-based cervical cytology slides were scanned using the FocalPoint(R) System. Ten black-and-white images that had the greatest probability of containing abnormality were acquired from each of 32 reference slides (16 negative samples, 3 samples of atypical squamous cells of uncertain significance, 5 samples of low-grade squamous intraepithelial lesions [LSIL], 5 samples of high-grade squamous intraepithelial lesions [HSIL], 1 adenocarcinoma in situ sample, and 2 carcinoma samples) and were transmitted as e-mail attachments in JPEG format to remote reading stations. The slides were interpreted independently by two pathologists and were assigned to either of two groups: 1) suspicious for >or=HSIL or 2) <or=LSIL. The interpretations were compared with the reference diagnoses. The specimens were then randomized, and the image sets were redistributed to the pathologists for another round of interpretation and scoring.

RESULTS

The initial and subsequent trials yielded similar results. Pooling the interpretations of the two pathologists, the concordance rate between reference and assigned diagnostic groups for each of the two trials was 84%, the false-positive rate was 8.3%, and the false-negative rate was 37.5%. Review of the discrepant slides revealed subtle cellular changes that may be utilized to reduce errors and, with training, to optimize sensitivity and specificity.

CONCLUSIONS

This procedure showed promise for allowing remote interpretation of device-selected images. The procedure may represent an effective way to centralize cervical cytology services and to allow the provision of services to previously unscreened populations that lack an effective cytology infrastructure.

FocalPoint slide classification algorithms show robust performance in classification of high-grade lesions on SurePath liquid-based cervical cytology slides

The FocalPoint Primary Screening System (FPPS) operates by assigning scores to each slide relative to the probability that an abnormality is present. This information ranks each slide within five "quintiles" (1 = highest risk, 5 = lowest risk) of the "review" population, allowing examining cytologists to understand the risk inherent in each slide. Such information helps to make the primary and quality control rescreening processes most efficient. This study examines the efficiency of AutoPap scoring and, thus, stratification of high-grade cases within a clinical trial setting. A total of 1,275 SurePath (TriPath, Burlington, NC) slides were screened on the FPPS. There were 124 high-grade cases in the set (32 HSIL, 5 AIS, and 87 cancers) as determined by cytologic truth adjudication. The efficiency of FPPS ranking was determined by analysis of the numbers of high-grade cases ranked into quintiles 1 (top 20%) and 1+2 (top 40%). FPPS places cases scored as "unsatisfactory" (HSIL, 3; Cancer, 18) into quintile 5 to ensure a manual review. These cases were excluded from the analysis. Overall, 58% of high-grade slides were classified as Q1 and 83% were classified as Q1+Q2. For HSIL, 66% were classified as Q1 and 90% as Q1+Q2. For Cancer, 59% were classified as Q1 and 84% as Q1+Q2. No high-grade slides were ranked in the lower "no review" population (slides that would receive no manual examination). The results confirm the robust performance of the FPPS classification algorithm. Far more high-grade slides are classified into the top two quintiles than would be expected by random chance alone (20% and 40%, respectively). These results validate the safety and effectiveness of this device on SurePath liquid-based slides.

A feasibility study of the use of the AutoPap screening system as a primary screening and location-guided rescreening device

BACKGROUND

The AutoPap 300 QC system (Tripath Imaging, Inc., Burlington, NC) is an automated device that was designed to screen conventionally prepared cervical smears and, more recently, thin-layer slide preparations. The system has been tested in large clinical trials.

METHODS

A total of 14,145 cervical smears obtained from participants in the Florence screening program were eligible for the study. Smears were processed first with the AutoPap system and were classified into three different categories: 1) no further review (NFR), 2) review, and 3) process review (PR). Conventional manual reading was performed by 10 experienced cytopathologists.

RESULTS

After AutoPap processing, 2398 smears were classified as NFR (16.9%), and 1818 smears were classified as PR (12.8%). Overall, there were 188 inadequate smears (1.3%) at conventional review and 125 inadequate smears (0.88%) at AutoPap review. Six-month repeat smears were prompted by 330 conventional reviews (2.3%) and by 222 AutoPap reviews (1.56%). Similarly, referral to colposcopy was prompted by 179 conventional reviews (1.2%) and by 147 AutoPap reviews (1.0%). Overall, 32 patients were diagnosed with high-grade cervical intraepithelial neoplasia as a result of assessment. Conventional reading detected 31 patients (28 patients were referred for colposcopy, and 3 patients were referred for repeat cytology), and the AutoPap system detected 30 patients (27 patients were referred for colposcopy, and 3 patients were referred for repeat cytology).

CONCLUSIONS

The current experience suggested that conventional reading and AutoPap reading of cervical smears had essentially the same sensitivity, with slightly greater specificity for the AutoPap system. Thus, comparisons of the AutoPap system and conventional reading should focus mainly on cost analysis.

Endometrial cells and the AutoPap System for primary screening of cervicovaginal Pap smears

In 1998, the AutoPap 300 received FDA approval for primary screening of conventional cervical smears. As approved, smears categorized as "no further review" and comprising up to 25% of the smears screened by the AutoPap 300 can be reported as negative for malignant and dysplastic cells without screening by a cytotechnologist. We studied 106 conventional cervical smears in which glandular endometrial cells had been identified by manual screening to assess the ability of the AutoPap System (TriPath Imaging, Inc., Burlington, NC) to (1) designate conventional Papanicolaou smears that contain endometrial cells for "review," and (2) stratify smears that contain endometrial cells as more or less likely to be abnormal. Although the number of cases in our study was small, our findings indicate that (1) the AutoPap System is slightly less sensitive than manual screening by experienced cytotechnologists for the detection of endometrial cells in conventional smears, as the System designated for "review" 94.3% of all smears containing endometrial cells, 92.9% of smears reported as atypical glandular cells of undetermined significance (AGUS) or endometrial adenocarcinoma, and 100% of the four smears with subsequently confirmed endometrial adenocarcinomas, (2) ranking of smears into quintiles by the AutoPap System did not provide additional diagnostically useful information with respect to endometrial pathology, (3) the number of endometrial cells in the smears did not correlate with quintile assignment, and (4) for most patients, routine use of the AutoPap System for primary screening of conventional cervical smears is unlikely to contribute to delay in the diagnosis of clinically significant endometrial lesions.

Prevention in the year 2002: some news, some issues

As we move into the new century, we continue to focus on ways to prevent disease, promote health, and prolong life. We have made enormous strides in some arenas, including deaths from infectious diseases caused by mass immunization, improved sanitation, better nutrition, heart disease caused by risk factor modification, and some cancers caused by smoking cessation and screening for early detection. Although we have made progress, many challenges remain.

Automated cervical cytology: meta-analyses of the performance of the AutoPap 300 QC System

The objective of this study was to review current knowledge regarding the performance of the AutoPap 300 QC System (NeoPath Inc., Redmond, WA) for automated cervical cytology screening. To this goal, we identified all studies published in the English language that included the AutoPap 300 QC automated cervical cytology system. The studies were obtained from a MEDLINE search through October 1998; additional sources were identified through cross-referencing. Studies concerning the AutoPap 300 QC System containing complete data are presented descriptively. Meta-analyses about the performance of the AutoPap 300 QC System were performed. The central goal of the meta-analyses was to estimate the overall false-negative rate of the AutoPap 300 QC System when applied in either of the two following modalities: primary screening and quality control. Of the 14 studies concerning the performance of the AutoPap 300 QC System as a primary screening modality, four studies provided complete data about the number of abnormal slides, review rate, and number of slides selected. Meta-analysis of these four studies indicate sensitivities ranging between 85 and 100 percent. Regarding the performance of the AutoPap 300 QC System in the quality control modality, of the 14 studies reviewed, 5 studies provided complete data including the number of false-negatives, review rate, and number of slides selected. Meta-analysis of these five studies indicate an average sensitivity of the AutoPap 300 QC System applied as a rescreening modality of 37 percent (95% CI; 34-40 percent), with observed salvage ratios of between 3.5 and 5.6 when review rate of the AutoPap 300 QC System was set at 10 percent.