PAPNET™ Cytological Screening System
Transcription
PAPNET™ Cytological Screening System
NEW INSTRUMENTATION PREVIEW PAPNET™ Cytological Screening System least 20 years. Automation offered the hope of efficiently and accurately processing the voluminous and labor-intensive nature of Pap smear screening. However, all previous attempts to automate the analysis of cervical smears have relied on classical algorithmic processing techniques. These techniques depend on finding the boundaries of objects and deriving simple morphological features (eg, area and density). While these techniques work well with simple objects in uncomplicated scenes, they are incapable of handling the complex and infinitely variable combinations of overlapping material typically found on Pap smears. Such attempts at Pap smear automation, therefore, have required some type of monolayer cell preparation. In order Input \' \1 ]/ J1 ~ >> The literature is replete with reports of failures of the cytology laboratory to diagnose invasive cancer. One study reported a cumulative falsenegative error rate for invasive cancer of about 50%. In the same study, the rate of screening errors for precancerous lesions was at least 28% . . . . Regardless of the percentages, it is clear that the error rate of cytologic screening for precancerous lesions and invasive cancer of the uterine cervix is quite substantial (Koss LG: The Papanicolaou Test for Cervical Cancer Detection: A Triumph and a Tragedy. JAMA 1989; 261:737-743). Attempts to bring automation to the analysis of cervical smears date back at to make a thin, monolayer preparation, only a subpopulation of all the cells obtained from the patient is used. Monolayer preparations thus represent a source of false-negative results due to the cell loss associated with subsampling and filtration or centrifugation. The very slides that typically cause laboratory screening false-negative results have very few abnormal cells on them; the loss of cells from the artificial processing and subsampling associated with all monolayer preparations increases the potential for falsenegative screening errors. Another disadvantage of monolayers is that they require a change in standard gynecology and pathology practice. Use of the conventionally prepared Pap smear as- J' \ Layers of ~j processing / elements i __-;?=» Connections ' L~ Output Fig 1. Neural networks are reminiscent of neurobiology not only in the sense that they learn from experience but also because they are characterized by highly parallel architectures composed of densely interconnected but relatively simple processing elements. 276 Laboratory Medicine Vol. 22, No. 4 April 1991 Downloaded from http://labmed.oxfordjournals.org/ by guest on October 26, 2016 PAPNET, a semiautomated cytological screening system, is available for investigational use from Neuromedical Systems, Inc (NSI®), of Suffern, NY. PAPNET is a computerized imaging system that uses neural network-emulating software to help detect abnormal cells on conventionally prepared and stained cervical Pap smears. NSI is the first company to apply this advanced, neural network computer technology to cytology. The PAPNET system has a false-negative rate of less than 3% (significantly less than that in the average laboratory), according to the initial clinical trial. The current manual practice of screening cervical smears has an intrinsic and unavoidable false-negative rate. This occurs primarily because relatively few abnormal cells are present among the vast amounts of normal material viewed daily. Essentially, the cytologist is searching for a "needle in a haystack." In any screening task, be it proofreading or Pap smear screening, whenever the vast majority of objects examined are not of concern (normal), psychological habituation takes place, which makes the proverbial needle that much harder to find. Sometimes it is missed altogether: traepithelial lesions or malignancies of other uterine sources. sures that the classical visual presentation of cytopathology (tumor diathesis, infective background) is maintained and intact and that no change to current medical practice is required. Operation Principle The PAPNET system utilizes both algorithmic and neural network computer technologies. Neural networks are parallel information processors that mimic the neurological ability to learn from experience. They, therefore, excel at recognizing subtle, hard-to-define patterns. This contrasts them from classical computers that require a step-bystep description (or algorithm) of how to distinguish one object from another. Neural networks have been used primarily in aerospace and defense to identify ambiguous objects (are they missiles or a flock of birds?). Because PAPNET includes a neural network (Fig 1), it can mimic the human cytotechnologist's ability to distinguish single or overlapping abnormal cells from the overlapping normal cells, debris, lymphocytes, blood, and neutrophils found on the conventionally prepared Pap smear. A robotic arm delivers conventionally prepared Pap smear slides from a preloaded slide cassette holder to an automated microscope stage. A bar code reader confirms patient identification. Slides are first scanned using a low power objective to locate cellular material and are then scanned using high power objectives with automatic focusing. A high-resolution, full-color video camera passes the images on to the primary algorithmic classifier. The primary classifier locates cell nuclei and other matter using morphological criteria. While most of the normal-appearing cells are screened out, approximately 1,000 to 10,000 potentially suspect objects (mostly overlapping normal cells) are passed to a neural network-based secondary classifier for evaluation. The neural network then classifies these images by generalizing from its training and selects 64 objects from each slide. These images are then stored on a data storage cartridge. Following the completion of PAPNET's run of 100 slides, the cytologist removes the data storage cartridge from the Scanning Station and inserts it into the Review Station. He or she now evaluates the 64 suspect objects from each slide selected by the PAPNET system on a high-definition, full-color monitor (Fig 2). In approximately 30 seconds, the cytologist can either confirm a negative diagnosis or determine that the case needs closer scrutiny, based on the presence of cells indicative of squamous in- An automated microscope with several components has a robotic arm that delivers slides from a preloaded cassette, which holds up to 100 slides, to the motorized stage. A bar-code reader confirms patient identification. A highresolution, full-color video camera obtains electronic images from a linear array of three objectives to provide magnification of 50x, 200x, and 400x. A slide dotter uses an ink-jet pen to mark the suspicious cells as specified by the cytologist during the follow-up review process. A central processing unit is the main controller of the Scanning system. It is a very high-speed computer system that includes a video display controller, a video monitor, a keyboard, a floppy disk drive, a large hard disk drive, and parallel and serial ports. Downloaded from http://labmed.oxfordjournals.org/ by guest on October 26, 2016 Fig 2. The PAPNET ™ System Review Station. System Description A minimum PAPNET configuration consists of one 4x3-ft Scanning Station and one 4x3-ft Review Station. This basic configuration has an approximate throughput of 40,000 to 50,000 slides per year. Each additional 40,000 slides per year would require an additional 4x3-ft Scanning Station. The scanning and review stations can be installed in different sections of the laboratory and are operated independently of each other. The basic power requirements are one 220-V 15-A outlet and one 110-V 15-A outlet for each Scanning Station and one 110-V 15-A outlet for each Review Station. The Scanning Station automatically scans Pap smear slides for atypical cells. It runs continuously and virtually unattended, storing 64 images from each slide for later inspection by the cytotechnologist on the Review Station. The Scanning Station consists of the following subsystems. The primary image processor performs several tasks. It finds areas of interest on the slides, it finds focus, it is used as the primary screening process for atypical cell detection, and it is used to compose color images for creating the 64-cell display grid. The neurocomputer is used as a pattern-recognition device in the secondary screening process for atypical cells. A data storage device with removable media is used to store the 64-cell images from each slide for later review. The images are stored in a digital data format with all of the images from 100 slides stored on a single small disk or tape cartridge. Uboratory Medicine Vol. 22, No. 4 April 1991 277 The Review Station is used for viewing the 64cell images selected for each slide by the Scanning Station. The Review Station is the workstation for the cytologist, and consists of the following subsystems: A robotic arm delivers slides from a preloaded cassette cartridge to an automated stage. The host CPU is the main controller of the Review system workstation. It is a very highspeed computer along with a VGA video display controller, a video monitor, a keyboard, a floppy disk drive, a large hard disk drive, and parallel and serial ports. The high-resolution RGB monitor is used to present the 64-cell images from the high-resolution display controller to the cytologist for review. This is the primary display screen to the operator. A bar code reader confirms patient identification. 000 Slides are scanned using a high resolution, color video camera with automatic focusing. A mouse is used as a pointing and selecting device by the cytologist. Basic Operation The PAPNET workflow is outlined in Fig 3. The PAPNET system is designed to use standard Papanicolaou-stained smears on glass slides. The smears are made and prepared just as in current laboratory practice. Monolayer preparation or Feulgen staining, two techniques that have been used in the past for automated cytology systems, are not necessary with PAPNET, although PAPNET can read slides prepared in these ways. A primary, "algorithmic" classifier locates cell nuclei and other objects using morphologic criteria. Cell nuclei and other objects are passed to a neural network based secondary classifier. Scanning Station A laboratory technician places the loaded slide holder cassette on the microscope elevator. He or she flips a toggle switch to turn on the vacuum that keeps the slide holder cassette in place. The technician then inserts an empty data cartridge into the data storage device. Then, using the system keyboard, the scanning is started. The rest of the operations are automatically performed by the Scanning software. The software aligns the stage, the objectives, and the robotic system by issuing commands to send each to its home position. Then the data cartridge is initialized. A directory is made, using the current date as directory name, and transferred onto the data cartridge. ZZZ3 64 suspect cells from every slide are digitally recorded for display and confirmation by a cytologist. Confirmed abnormal cells are then separately recorded and "dotted" on the glass slide by an automated marker. A report is generated summarizing findings by the PAPNET system. Fig 3. The PAPNET workflow outline. 278 Laboratory Medicine Vol. 22, No. 4 April 1991 Downloaded from http://labmed.oxfordjournals.org/ by guest on October 26, 2016 A data storage device with removable media is used to retrieve the 64-cell images from each slide as stored by the Scanning Station and to store the locations of any cells that are tagged by the cytologist for dotting during the review process. PAPNET used: for Rescreening Pathologist 1 False negative Cytotechs 98 PAPNET 97*\| 100 Slides *False negative rate <19< CD \ en PAPNET used: for Prescreening Pathologist Examination Positive report to - |,GYN within -^24 hours ?! PAPNET 97 Cytotech Examination 97 • \ i 100 Slides *False negative rate <196 CD CD Fig 5. Hypothetical case. With a more rapid turnaround time on positive results, a false-negative result due to lab screening error could only occur if a positive result is missed by the the manual and PAPNETassisted examinations. The robotic system then picks up the first slide and places it onto the microscope stage, at which point the bar code is read. The slide is then scanned. When the scan of the first slide is completed, the grid of 64-cell images is stored on the data cartridge and assigned the bar code number as its file name. The remaining slides in the cassette are processed in the same way. The processing of all 100 slides in a cassette should take, on average, about 16 hours. Should a technician start a scanning run at 4 PM, the images from these 100 slides would be available for cytologist review around 8 AM the next morning. Should a laboratory require additional throughput, NSI installs additional Scanning Stations that are run in parallel. After the 100 slides are processed, the system stops running automatically, indicating on the operator's screen that the run is finished. The technician then flips the toggle switch, releasing and removing the slide holder cassette. He or she ejects the data cartridge from the data storage device and places it on the back of the corresponding slide holder cassette. The technician may then start another run as described above. Downloaded from http://labmed.oxfordjournals.org/ by guest on October 26, 2016 Fig 4. Hypothetical case. A false-negative result due to lab screening error could only occur if a positive case is missed by both the manual and PAPNET-assisted examinations. Review Station The cytologist removes the data cartridge from the Scanning Station's data storage device and inserts it into the Review Station's data storage device. Then the cytologist selects the appropriate menu command to start the Review function. The Review function retrieves the first 64-cell image grid from the first slide and loads it onto the high-resolution RGB monitor. The cytologist can now review the slide. The 64-cell images are presented (at 200x) as a single 8x8 grid of "tiles" on the screen and also as four screens of 4x4 tiles. Each tile presents the suspicious object in its center with a 128x104-um contextual surround. The cytologist uses a mouse as a pointing and selecting device and can zoom in the image for greater detail (400x) and zoom it back out as required using one of the mouse buttons. When the cytologist detects an abnormal cell, the arrow cursor is placed in the subject tile and the other mouse button is pressed. This places a red border around the subject tile, facilitating later review of the case on the video screen. In addition, depressing this mouse button will place a dot of ink on the glass microscope slide at the location of the selected cell. Any number of tiles on each slide can be tagged. After reviewing all 64 tiles, the cytologist will classify the case. Then, the Review function retrieves the next 64cell image grid and loads it onto the high-resolution monitor. The cytologist can then classify the next case as described above. When review of all 100 slides is completed, the slide holder cassette is placed back into the microscope. The data cartridge is removed from the Review Station and is inserted into the data storage device on the Scanning Station. The CPU reads the review log file and determines whether any slides have been flagged by the cytologist for dotting during the review process. If there are any, the robotic arm retrieves them from the appropriate slide holder cassette position. The bar code is read to verify that it is the correct slide. Then the automated stage moves the slide to the positions recorded by the cytologist during the re- Laboratory Medicine Vol. 22, No. 4 April 1991 279 Conclusion By optimizing the combination of machine vision and human intelligence, PAPNET can be currently used on an investigational basis in one of two quality assurance modes: rescreening and prescreening. In the same amount of time that it now takes to perform the 280 Laboratory Medicine Vol. 22, No. 4 April1991 widely practiced 10% random rescreen of negative smears, a cytologist using PAPNET can perform a 100% rescreen or prescreen. Using PAPNET as a rescreener, a false-negative result caused by a laboratory screening error could only occur in the unlikely event that a positive case is missed by both the manual and the PAPNET-assisted examination. Prescreening with PAPNET would offer the additional benefit of a quicker turnaround time on positive cases, since abnormal cases flagged by the PAPNET system would receive the priority attention of the pathologist. Figures 4 and 5 are hypothetical cases that illustrate how PAPNET could be implemented. (Data are based on initial clinical trials.) A limited number of PAPNET systems are available for investigational use pending FDA clearance for clinical use—Robert P. De Cresce, MD, MBA, and Mark 5. Lifshitz, MCU Downloaded from http://labmed.oxfordjournals.org/ by guest on October 26, 2016 Quality Control Quality Control software runs diagnostic routines to continually assess the working status of the system. To ensure the quality of the system/user interface, the grids of 64 tiles can be reviewed no more rapidly than at a preset rate. In addition, the cytologist will be required to use the mouse to drag the cursor through the center of each review tile. This actually audits the fact that the cytologist visually attends to each suspicious cell. the PAPNET™ cytological screening system semiautomates the Pap smear screening process in a commercially viable way. One important advantage of the PAPNET system over other attempts at automated cytology is its ability to process conventionally prepared and stained cervical smears. Thus, the currently used methods of sample collection and processing can be maintained without introducing new sample collection and preparation techniques that would deviate from standard practice, be difficult to implement on a large scale, and cause unnecessary loss of diagnostic cells and conventional diagnostic cues. view process and an automatic ink dotter puts a dot on the glass.